fix(plex): advertise the plex.direct name, not the bare VIP (server vanished from the plex.tv account)

customConnections held two hand-typed URLs and neither could ever complete, so Plex had nothing publishable and the server disappeared from the account: - https://192.168.2.24:32400 -- the server's cert is *.130072f20c334d6c8f63d0dfef87309b.plex.direct with NO IP SAN, so bare-IP https fails TLS for every client. Measured: a continuous storm of "CERT: incomplete TLS handshake ... certificate unknown" from real LAN clients arriving through the plex-l4-tcp VIP. - https://plex.home.keithjackson.dev:32400 -- nothing listens; the Gateway serves that host on 443 (measured http=000 against 192.168.2.20:32400). With no completable connection plex.tv published nothing ("Mapped - Not Published"), Plex fell back to Relay, and clients reported "not authorized". Fix: advertise <dashed-ip>.<serverhash>.plex.direct -- public DNS that resolves to the embedded address, so it points at the LAN VIP and matches the wildcard cert. Verified with full TLS validation (no curl -k): 200 + verify OK, where the bare-IP form fails to connect. Declared as ADVERTISE_IP so git owns the value instead of a web form: the image's 40-plex-first-run writes it into customConnections on every pod start, outside the claim guard. Does NOT fix off-LAN access -- the advertised address is RFC1918 and the WAN tcp/32400 forward remains unverified. ManualPortMappingMode=1 with an absent port is left as-is (no GitOps path exists for it). Touches clusters/** -> sensitive paths -> needs the operator's /approve-override, deliberately not self-posted (CLAUDE.md 1). Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Merge pull request 'fix(cilium): point cgroup.hostRoot at the host cgroup2 root (socket-LB never covered host processes)' (#42 ) from fix/cilium-cgroup-hostroot into main
2026-10-04 02:53:01 -05:00 · 2026-10-04 07:10:31 +00:00 · 2026-10-04 07:08:19 +00:00 · 2026-10-04 02:03:42 -05:00 · 2026-10-04 06:50:56 +00:00 · 2026-10-04 01:34:52 -05:00
94 changed files with 10801 additions and 149 deletions
@@ -36,6 +36,36 @@ sensitive path, so the policy cannot be widened or narrowed without a human.

 That stage is **fail-closed** — if it cannot evaluate, it blocks. Unlike `gitea-settings`, where a
 missing credential is a lock-out, here a skip would be a bypass.
+
+**The administrator override (2026-10-01).** The rule above was unsatisfiable for the one human
+who runs this repo: Gitea refuses to let an author approve their own PR, and `jakman85` is the
+only collaborator — so a sensitive-path PR the operator opened could never be merged, *including
+the PR that would fix that.* An administrator may therefore waive the review by posting this on
+the PR:
+
+    /approve-override <reason>
+
+**Post it as a PR *review* of type "Comment", not in the plain comment box.** CI's
+`SETTINGS_API_TOKEN` is minted `[read:repository, read:admin]`, and `/issues/{n}/comments` sits
+behind Gitea's *issue* scope — so an override typed into the comment box is **invisible to CI**,
+and the stage reports it as a missing approval rather than as a permission problem (measured on
+PR #38, 2026-10-02). `/pulls/{n}/reviews` is readable, so a review body works today. Both sources
+are accepted and the gate names which it used; the comment box starts working once
+`infra/gitea/set-ci-token.sh` is re-run to pick up `read:issue`.
+
+The directive must **start a line**, the **reason is mandatory**, the poster's `admin`/`owner`
+permission is re-checked **live** when the gate runs, and it must be **newer than the
+head commit** — push again and you must override again. The overrider, the timestamp and the
+reason are printed in the gate log and the job summary. For an administrator, "a change to the
+gate cannot be gated by the gate" is therefore **advisory, not structural**; what is kept is an
+attributed record of each waiver. Note that the agent account `gitea-admin` also holds `owner`
+here, so an agent *can* clear this on its own PR — **do not**: posting that directive to unblock
+your own agent work is exactly the self-approval the stage exists to prevent. Get the operator
+to post it, or move the change out of the sensitive paths.
+
+**`required_approvals` on `main` is `0`**, not 1 — the gate is the approver. That drops the
+approval *count* only: `block_admin_merge_override` and `enable_status_check` both remain `true`,
+so **a red gate is unmergeable by everyone, administrators included.**
 - **Do not `kubectl apply` / `edit` / `patch` a durable change.** Flux reconciles `main`;
  anything applied by hand is drift that the next reconcile reverts or, worse, adopts.

@@ -53,12 +83,19 @@ Verified against the live Gitea on 2026-09-29:
 | `main` protected, direct push rejected | ✅ `protected: true`, `user_can_push: false` — **verified by a real rejected push**, not by the API's word |
 | PR blocked until the gate passes | ✅ `enable_status_check: true`, context `gate / gate (pull_request)` |
 | Gate runs on every PR | ✅ runner `incluster-dind` online; ~1m35s |
-| Approval required before merge | ✅ `required_approvals: 1` |
+| Gate cannot be overridden, by anyone | ✅ `block_admin_merge_override: true` |
+| Human approval required before merge | ⬜ **No — `required_approvals: 0` since 2026-10-01.** The gate is the approver; sensitive paths are gated by `ci/approval-policy.sh` instead. See §1 |

-All four were FALSE as recently as 2026-09-29, while this repo's own log asserted they were
+The first three were FALSE as recently as 2026-09-29, while this repo's own log asserted they were
 true. They are true now because each was measured, and the push rejection was provoked rather
 than assumed. **Re-measure before relying on any of them** — see §2b.

+The fourth row deliberately changed on 2026-10-01. `required_approvals: 1` was a **lock-out**, not
+a protection: with a single collaborator and Gitea refusing author self-approval, nothing could
+satisfy it (measured on PR #35 — green gate, zero reviews, unmergeable). It also contradicted the
+model already written into `ci/approval-policy.sh` and `ci/gate.sh`. Rationale and the live-change
+script: `infra/gitea/allow-admin-self-merge.sh`.
+
 Still not covered: `vars.CLUSTER_AVAILABLE` is unset, so **L2/L3 conformance does not run in
 CI**. A green PR therefore proves the L0/L1 offline subset, not runtime behaviour.

@@ -100,11 +100,15 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://jackett.${BASE_DOMAIN}/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --api-route=^/api/
+        - --cookie-name=_oauth2_proxy_jackett
+        - --cookie-domain=jackett.${BASE_DOMAIN}
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -125,6 +129,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -14,10 +14,12 @@ resources:
 - nextcloud-db-backup-credentials-prod.sops.yaml
 - nextcloud-db-backup-credentials.sops.yaml
 - nextcloud-db.yaml
+- pihole-bootstrap-secret.sops.yaml
 - pihole-dns.yaml
 - pihole-oauth2-proxy-credentials.sops.yaml
 - pihole.yaml
 - plex-claim-secret.sops.yaml
+- plex-remote.yaml
 - plex.yaml
 - qbittorrent-bootstrap-secret.sops.yaml
 - qbittorrent-oauth2-proxy-credentials.sops.yaml
@@ -0,0 +1,72 @@
+# =================================================================================================
+# PiHole WEB/API ADMIN PASSWORD — the one piece of PiHole state that was NOT declared anywhere.
+# =================================================================================================
+# WHY THIS FILE EXISTS. Measured on live prod 2026-10-04, `pihole.toml` marks every non-default
+# setting with `### CHANGED`, and names the ones that came from the environment `### CHANGED (env)`.
+# The complete list was:
+#
+#     dns.upstreams          ### CHANGED (env)   <- declared, catalog/pihole.yaml
+#     dns.listeningMode      ### CHANGED (env)   <- declared, catalog/pihole.yaml
+#     dns.rateLimit.count    ### CHANGED (env)   <- declared, catalog/pihole.yaml
+#     misc.dnsmasq_lines     ### CHANGED (env)   <- declared, catalog/pihole.yaml
+#     files.macvendor        ### CHANGED         <- set by the image itself, container path
+#     webserver.api.pwhash   ### CHANGED         <- SET OUT OF BAND. Declared NOWHERE. This file.
+#
+# So exactly ONE operator-set value lived only on the Longhorn PVC: the admin password, as a
+# one-way BALLOON-SHA256 hash. Nobody could recover it (a hash does not reverse), nothing in git
+# produced it, and a rebuild from git would have come up with a DIFFERENT password — or none.
+#
+# THE REASON THIS BLOCKS HA, which is why it is fixed first. The next step is running more than one
+# PiHole replica, and a replica with a fresh PVC would come up with its own unrelated password. The
+# admin UI would then accept or reject the same credential depending on which pod answered. Declaring
+# it here makes every replica, and every rebuild, identical by construction.
+#
+# MECHANISM: `webserver.api.password` is a real, settable FTL config key — verified in the running
+# image, where `pihole setpassword` is itself just `setFTLConfigValue "webserver.api.password"` with
+# the comment "pihole-FTL will automatically hash the password". So FTL takes CLEARTEXT on this key
+# and writes the hash into `webserver.api.pwhash` itself. catalog/pihole.yaml therefore passes this
+# value as `FTLCONF_webserver_api_password`, and the hash on the PVC becomes a derived artefact
+# rather than the source of truth.
+#
+# DEFENCE IN DEPTH, NOT THE ONLY GATE. The admin UI is already behind the oauth2-proxy sidecar
+# (`--allowed-group=platform-admins`), so this password is a SECOND gate. It is kept rather than set
+# empty because the `pihole` Service still publishes port 80 (PiHole direct, bypassing the proxy on
+# :4180) and only the CiliumNetworkPolicy closes it — and the default-deny policy in this repo has a
+# documented drift history. An unauthenticated admin UI that can rewrite LAN DNS is not a good thing
+# to leave resting on one policy staying correct. Narrowing that Service port is tracked follow-up.
+#
+# ROTATION: `sops` this file in your own terminal, change the value, commit. Flux re-applies the
+# Secret, the pod restarts, FTL re-hashes. Do NOT `sops -d` it into a shared terminal or a log
+# (CLAUDE.md section 6); read it in an editor. The value below was generated randomly at authoring
+# time and has never been printed to a transcript.
+apiVersion: v1
+kind: Secret
+metadata:
+    name: pihole-bootstrap
+    namespace: pihole
+    labels:
+        app.kubernetes.io/name: pihole
+type: Opaque
+stringData:
+    #ENC[AES256_GCM,data:TMOuRyJeb5IVn3lZPXY7lgh1KGxu7GY9+WIOLHpMB50JqzmM2tHIaWUWKTt1JtZoX/E5/Y7WCUTnfkJLhRBN2zV2MwWRBSLYVy5CMg7xV8PTLRZJYAVUYS5Trg==,iv:q+DpfT6PQfY5ZuNsmP7Ddl+PDLjSm89fJICAhYoW64k=,tag:axijaO3DaxIZgZ3LeK2hXA==,type:comment]
+    webui-password: ENC[AES256_GCM,data:7FcK/bUW79M46f/MlouAmATKbMnoUTNuMUjtqJ5vpU4=,iv:I8DBFB39uXZV42Ie2IPwSefmMgdKEtiDknfc9Tu7h5E=,tag:L423784UrDU3TvylpHc6KQ==,type:str]
+sops:
+    kms: []
+    gcp_kms: []
+    azure_kv: []
+    hc_vault: []
+    age:
+        - recipient: age175l8mdpdfu7hq0aml3exs3gykarwjmuvr23sq56s3wjejnyk44wsp4mfzg
+          enc: |
+            -----BEGIN AGE ENCRYPTED FILE-----
+            YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNTUxOSB0MkRSNjZLZlpXYkcvaVJL
+            cjlqb2wrcWREZnZLdGpsTHNPWEpseHlRZnpjCnJVM1FDQ2M2dXlaS1JBZ2k2V1BP
+            Yi9QUDZSMzZsL09pTi93RTdjdGxtY2MKLS0tIDIrQTZIVWdhMGIxSCtlUFN2Mlho
+            SWVybld4VjJyaHRSKzZZU0xsUUV2SEEKbDYHV/qBZVspK7DpBW3vYjRMJ/HmAN+9
+            G9bOlNEp0jAtcNZEn6lpQ9ORgvmIrhJHoBQRLabdMWzNWSZsNn56tA==
+            -----END AGE ENCRYPTED FILE-----
+    lastmodified: "2026-10-04T06:01:27Z"
+    mac: ENC[AES256_GCM,data:rhTMTOwLOtHXHgqI5M2ZW/6w+UpLpxBlTk7NoN0wmGSqOcE+OGrXA1DhkZrJweyQT0RSNSSBm/6QMztLP81eepOzdp2rEkwvbQPK/l6K9g16Ww9c1IatS3NQsKEB8AoMDx7jBQ6exXYgggPx3MJsNYbE+WyloQ96U9KPdrecAM8=,iv:n34v2zX9D5xsaR1PFCH94mNMiuRfhUyI2lseUCoqiTA=,tag:9YmN9nIVLXfsGYIb0bUl2w==,type:str]
+    pgp: []
+    encrypted_regex: ^(data|stringData)$
+    version: 3.9.4
@@ -35,34 +35,18 @@ spec:
    - protocol: TCP
      port: 53
 ---
-# PVC pihole-config (R-STATE)
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: pihole-config
-  namespace: pihole
-  labels:
-    app.kubernetes.io/name: pihole
-spec:
-  accessModes:
-  - ReadWriteOnce
-  storageClassName: longhorn
-  resources:
-    requests:
-      storage: 2Gi
---
-# Deployment (R-PLAT-02)
+# StatefulSet (R-PLAT-02)
 apiVersion: apps/v1
-kind: Deployment
+kind: StatefulSet
 metadata:
  name: pihole
  namespace: pihole
  labels:
    app.kubernetes.io/name: pihole
 spec:
-  replicas: 1
-  strategy:
-    type: Recreate
+  replicas: 2
+  serviceName: pihole
+  podManagementPolicy: Parallel
  selector:
    matchLabels:
      app.kubernetes.io/name: pihole
@@ -90,9 +74,21 @@ spec:
          value: '0'
        - name: FTLCONF_misc_dnsmasq_lines
          value: address=/${BASE_DOMAIN}/${GATEWAY_VIP};address=/unifi.${BASE_DOMAIN}/${UNIFI_L4_VIP};address=/crafty-mc.${BASE_DOMAIN}/${CRAFTY_L4_VIP};address=/pihole-dns.${BASE_DOMAIN}/${PIHOLE_DNS_VIP}
+        - name: FTLCONF_webserver_api_password
+          valueFrom:
+            secretKeyRef:
+              name: pihole-bootstrap
+              key: webui-password
        volumeMounts:
        - name: config
-          mountPath: /data/config
+          mountPath: /etc/pihole
+        lifecycle:
+          postStart:
+            exec:
+              command:
+              - /bin/bash
+              - -c
+              - "# Assert the declared adlists in gravity.db. Idempotent: steady state writes nothing and\n# does NOT run a gravity rebuild.\nset -u\nGRAVITY=/etc/pihole/gravity.db\n\n# THE DECLARED SET \u2014 one URL per line, the source of truth for what PiHole blocks.\n# To add a list: add a line here, commit, let Flux reconcile. Nothing else to do.\nDECLARED='https://raw.githubusercontent.com/StevenBlack/hosts/master/hosts'\n\n# On a fresh PVC, FTL creates gravity.db a moment after the container starts. Wait, bounded,\n# and NEVER block or fail the start: a missing DB here must not crashloop DNS for the LAN.\ni=0\nwhile [ ! -s \"$GRAVITY\" ] && [ \"$i\" -lt 90 ]; do sleep 2; i=$(expr $i + 1); done\nif [ ! -s \"$GRAVITY\" ]; then\n  echo \"pihole-adlists: gravity.db absent after 180s - leaving adlists untouched\" >&2\n  exit 0\nfi\n\nsql() { pihole-FTL sqlite3 \"$GRAVITY\" \"$1\"; }\nCHANGED=0\n\n# 1) Every declared URL must be present AND enabled.\necho \"$DECLARED\" | while IFS= read -r url; do\n  [ -n \"$url\" ] || continue\n  printf '%s\\n' \"$url\"\ndone > /tmp/declared.txt\nwhile IFS= read -r url; do\n  [ -n \"$url\" ] || continue\n  # Single quotes inside a URL would break this SQL; adlist URLs do not contain them.\n  have=$(sql \"select count(*) from adlist where address='$url';\")\n  if [ \"$have\" = \"0\" ]; then\n    echo \"pihole-adlists: adding $url\"\n    sql \"insert into adlist (address,enabled,comment) values ('$url',1,'declared in catalog/pihole.yaml');\"\n    CHANGED=1\n  else\n    en=$(sql \"select enabled from adlist where address='$url';\")\n    if [ \"$en\" != \"1\" ]; then\n      echo \"pihole-adlists: re-enabling $url\"\n      sql \"update adlist set enabled=1 where address='$url';\"\n      CHANGED=1\n    fi\n  fi\ndone < /tmp/declared.txt\n\n# 2) Anything enabled in the DB but NOT declared gets disabled (never deleted - see above).\nsql \"select address from adlist where enabled=1;\" | while IFS= read -r addr; do\n  [ -n \"$addr\" ] || continue\n  if ! grep -qxF \"$addr\" /tmp/declared.txt; then\n    echo \"pihole-adlists: disabling undeclared $addr\"\n    sql \"update adlist set enabled=0 where address='$addr';\"\n    echo changed >> /tmp/adlist-changed\n  fi\ndone\n[ -s /tmp/adlist-changed ] && CHANGED=1\nrm -f /tmp/adlist-changed /tmp/declared.txt\n\n# 3) Rebuild gravity ONLY if the declared set actually moved.\nif [ \"$CHANGED\" = \"1\" ]; then\n  echo \"pihole-adlists: declared set changed - running gravity update\"\n  pihole -g || echo \"pihole-adlists: gravity update failed; existing gravity.db retained\" >&2\nelse\n  echo \"pihole-adlists: already matches the declared set - nothing to do\"\nfi\nexit 0\n"
      - name: oauth2-proxy
        image: quay.io/oauth2-proxy/oauth2-proxy:v7.8.1
        imagePullPolicy: IfNotPresent
@@ -108,11 +104,14 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://pihole.${BASE_DOMAIN}/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_pihole
+        - --cookie-domain=pihole.${BASE_DOMAIN}
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -133,6 +132,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -157,10 +159,39 @@ spec:
            port: http
          initialDelaySeconds: 5
          periodSeconds: 15
-      volumes:
-      - name: config
-        persistentVolumeClaim:
-          claimName: pihole-config
+      affinity:
+        podAntiAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+          - labelSelector:
+              matchLabels:
+                app.kubernetes.io/name: pihole
+            topologyKey: kubernetes.io/hostname
+  volumeClaimTemplates:
+  - metadata:
+      name: config
+      labels:
+        app.kubernetes.io/name: pihole
+    spec:
+      accessModes:
+      - ReadWriteOnce
+      storageClassName: longhorn
+      resources:
+        requests:
+          storage: 2Gi
+---
+# PodDisruptionBudget — keep one replica through a drain (availability.replicas)
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: pihole
+  namespace: pihole
+  labels:
+    app.kubernetes.io/name: pihole
+spec:
+  minAvailable: 1
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: pihole
 ---
 # Service
 apiVersion: v1
@@ -263,11 +294,11 @@ spec:
  - name: pihole.slis
    rules:
    - alert: pihole-down
-      expr: kube_deployment_status_replicas_available{namespace="pihole", deployment="pihole"} == 0 or absent(kube_deployment_status_replicas_available{namespace="pihole", deployment="pihole"})
+      expr: kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"} == 0 or absent(kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"})
      for: 5m
      labels:
        severity: critical
        service: pihole
      annotations:
        summary: pihole is down
-        description: 'pihole: its Deployment has no available/ready replica for 5m, or the signal is gone entirely (SLI: up, via kube_deployment_status_replicas_available).'
+        description: 'pihole: its StatefulSet has no available/ready replica for 5m, or the signal is gone entirely (SLI: up, via kube_statefulset_status_replicas_ready).'
@@ -0,0 +1,126 @@
+---
+# =================================================================================================
+# PLEX REMOTE ACCESS — raw tcp/32400 north-south exposure, and the policy that admits it.
+# =================================================================================================
+# HAND-WRITTEN, NOT RENDERED. A non-catalog app-tier unit anchored to the `plex` ServiceDefinition
+# and placed into wave 0 by `catalog/render.py --emit-waves` via its EXTRA_UNITS table — the same
+# mechanism gitea-actions and pihole-dns use. It is hand-written for the SAME reason those are:
+# schemas/servicedefinition.schema.json is `additionalProperties: false` and has NO field for a
+# type=LoadBalancer Service, for a pinned lbipam VIP, or for raw-TCP north-south exposure at all.
+# `ingressRoutes` renders HTTPRoutes on the shared Gateway, which is an L7 HTTPS terminator.
+#
+# WHY THE L7 GATEWAY IS NOT ENOUGH HERE — and it is NOT because the Gateway fails to serve Plex.
+# Measured 2026-10-03 through the Gateway VIP: `GET https://plex.home.keithjackson.dev/identity`
+# returned HTTP 200 with `claimed="1"` and the real machineIdentifier, and the protected endpoints
+# (/library/sections, /status/sessions, /:/prefs) all returned 401. The L7 path WORKS.
+#
+# The problem is DISCOVERY, not proxying. Plex apps (phone, TV, Roku) do not resolve
+# `plex.${BASE_DOMAIN}`; they ask plex.tv for the server's connection list. A server gets onto that
+# list one of two ways: it advertises a custom URL (`customConnections`, the ADVERTISE_IP env), or it
+# publishes a reachable public IP:port for its own tcp/32400. Measured in the live Preferences.xml on
+# 2026-10-03: `customConnections` ABSENT, `ManualPortMappingMode` ABSENT, while
+# `PublishServerOnPlexOnlineKey="1"` — the server was trying to publish and had nothing reachable to
+# publish, so remote clients saw NO connection at all. The operator chose the NATIVE path (this
+# file) over advertising the Gateway URL: Plex then terminates its own TLS with its
+# `*.plex.direct` certificate and video streams bypass the Envoy L7 hop entirely.
+#
+# VIP: ${PLEX_L4_VIP} (cluster-vars, R-ENV-02 — dev 192.168.1.57, prod 192.168.2.24), PINNED via
+# lbipam.cilium.io/ips out of the CiliumLoadBalancerIPPool `default-pool`
+# (${CILIUM_LB_POOL_START}..${CILIUM_LB_POOL_STOP}) and ARP-announced on the flat LAN by the existing
+# CiliumL2AnnouncementPolicy `default-l2` (platform-config/cilium/*). No new Cilium config is needed;
+# this Service just consumes the pool. PINNING IS NOT OPTIONAL: the address is typed into a WAN
+# port-forward on the SOHO router, which is hand-maintained and off-repo, so it can never move.
+# 192.168.2.24 verified FREE live 2026-10-03 — `default-pool` reported IPsTotal 11 / IPsUsed 4 /
+# IPsAvailable 7, and the only four allocations were .20 (cilium-gateway-shared), .21 (unifi-l4-*),
+# .22 (crafty-l4-*) and .23 (pihole-dns-*).
+#
+# ONE SERVICE, TCP ONLY. Unlike unifi/crafty/pihole-dns this needs no second Service and no
+# lbipam.cilium.io/sharing-key: Plex's remote access is tcp/32400 and nothing else. (Plex's
+# DLNA/GDM discovery uses UDP, but that is LAN-only multicast and is not part of remote access.)
+#
+# NO `loadBalancerSourceRanges`, DELIBERATELY — and this is the one place this file diverges from
+# its closest sibling. apps/wave-0/pihole-dns.yaml pins its source range to ${LAN_CIDR} because a
+# resolver must never answer the internet. This Service is the OPPOSITE case: admitting remote
+# clients from arbitrary internet addresses is its entire purpose, so there is no range to narrow
+# to. Plex authenticates every caller itself (measured above: 401 on every protected endpoint
+# without a token), which is what makes an unfiltered port defensible here.
+#
+# NO external-dns ANNOTATION, DELIBERATELY. unifi and crafty both carry
+# `external-dns.alpha.kubernetes.io/hostname` so external-dns publishes their VIP as a public
+# Cloudflare record. This Service must not, for pihole-dns's reason: ${PLEX_L4_VIP} is an RFC1918
+# address that is useless in public DNS. Remote clients get this server's address from plex.tv, not
+# from the zone. The existing public `plex.${BASE_DOMAIN}` record is published from the rendered
+# HTTPRoute via the gateway source and is NOT affected by this file — the web UI keeps working
+# through the Gateway exactly as before.
+#
+# -------------------------------------------------------------------------------------------------
+# externalTrafficPolicy: Cluster — inherited from crafty's 8-DAY LAN OUTAGE, not re-derived.
+# -------------------------------------------------------------------------------------------------
+# `Local` is the tempting choice because it preserves the client source IP, which this service
+# actually has a use for (see the allowedNetworks note below). It is still WRONG here, for the
+# reason platform/crafty/service-l4.yaml documents at length: Cilium's L2 announcer elects ONE lease
+# holder per Service and does NOT follow the backend pod, so with `Local` the VIP can be announced
+# by a node that has no Plex pod, which black-holes the port while every object still reports
+# healthy. Plex is a single-replica Deployment with `strategy: Recreate` pinned to wherever the
+# scheduler put it, so the mismatch is the normal case, not a rare one. Take the SNAT.
+#
+# KNOWN CONSEQUENCE OF THAT CHOICE, stated rather than discovered later: `Cluster` SNATs the client
+# to the ingress NODE's address (192.168.2.1x), which falls inside the `192.168.0.0/16` entry of
+# catalog/plex.yaml's ALLOWED_NETWORKS. Plex will therefore classify every REMOTE client as LAN.
+# Measured 2026-10-03 on the Gateway path, which SNATs the same way: Plex logged the request as
+# `[10.0.2.214:36040 (Allowed Network (WAN))]`. This is a METERING effect, not a security one —
+# authorization is unchanged (401s above), but Plex's remote-quality/bandwidth limits will not bind
+# on these streams because it believes they are local. Narrowing ALLOWED_NETWORKS cannot fix it
+# without also de-classifying genuine LAN clients, so it is accepted.
+apiVersion: v1
+kind: Service
+metadata:
+  name: plex-l4-tcp
+  namespace: plex
+  labels:
+    app.kubernetes.io/name: plex
+  annotations:
+    lbipam.cilium.io/ips: "${PLEX_L4_VIP}"
+spec:
+  type: LoadBalancer
+  externalTrafficPolicy: Cluster
+  selector:
+    app.kubernetes.io/name: plex
+  ports:
+    - name: plex-remote
+      port: 32400
+      targetPort: 32400
+      protocol: TCP
+---
+# INGRESS allow for the raw port (R-NET-01/02).
+#
+# The rendered plex.yaml already carries `plex-allow-gateway-ingress` (fromEntities: ingress ->
+# 32400) and three reciprocal allows for sonarr/radarr/seerr. CiliumNetworkPolicies are ADDITIVE, so
+# this file adds the off-cluster case without touching those.
+#
+# `world` is Cilium's entity for anything outside the cluster, which is what a remote Plex client is.
+# `host` and `cluster` mirror platform/crafty/networkpolicy.yaml's game-port rule and are what make
+# this work under `externalTrafficPolicy: Cluster`: after SNAT the packet's source is the ingress
+# NODE, which Cilium classifies as `host` (same node) or `remote-node` (a different one) — NOT
+# `world`. Listing `world` alone would compile cleanly and then drop every remote stream, so all
+# four entities are named.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: plex-allow-remote-32400
+  namespace: plex
+  labels:
+    app.kubernetes.io/name: plex
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: plex
+  ingress:
+    - fromEntities:
+        - world
+        - cluster
+        - host
+        - remote-node
+      toPorts:
+        - ports:
+            - { port: "32400", protocol: TCP }
@@ -137,6 +137,11 @@ spec:
            secretKeyRef:
              name: plex-claim
              key: PLEX_CLAIM
+        - name: ADVERTISE_IP
+          value: ${PLEX_ADVERTISE_URL}
+        resources:
+          limits:
+            gpu.intel.com/i915: 1
        volumeMounts:
        - name: config
          mountPath: /config
@@ -167,6 +172,28 @@ spec:
    port: 32400
    targetPort: 32400
 ---
+# CiliumNetworkPolicy reciprocal ingress allow (R-NET-02): plex-allow-radarr-32400
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: plex-allow-radarr-32400
+  namespace: plex
+  labels:
+    app.kubernetes.io/name: plex
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: plex
+  ingress:
+  - fromEndpoints:
+    - matchLabels:
+        app.kubernetes.io/name: radarr
+        k8s:io.kubernetes.pod.namespace: radarr
+    toPorts:
+    - ports:
+      - port: '32400'
+        protocol: TCP
+---
 # CiliumNetworkPolicy reciprocal ingress allow (R-NET-02): plex-allow-seerr-32400
 apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
@@ -189,6 +216,28 @@ spec:
      - port: '32400'
        protocol: TCP
 ---
+# CiliumNetworkPolicy reciprocal ingress allow (R-NET-02): plex-allow-sonarr-32400
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: plex-allow-sonarr-32400
+  namespace: plex
+  labels:
+    app.kubernetes.io/name: plex
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: plex
+  ingress:
+  - fromEndpoints:
+    - matchLabels:
+        app.kubernetes.io/name: sonarr
+        k8s:io.kubernetes.pod.namespace: sonarr
+    toPorts:
+    - ports:
+      - port: '32400'
+        protocol: TCP
+---
 # CiliumNetworkPolicy gateway-ingress allow (R-NET-01/02): plex-allow-gateway-ingress
 apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
@@ -171,11 +171,14 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://qbittorrent.${BASE_DOMAIN}/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_qbittorrent
+        - --cookie-domain=qbittorrent.${BASE_DOMAIN}
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -196,6 +199,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -1,6 +1,8 @@
 # promtool rule-unit tests for the wave-0 per-app `<app>-down` SLI rules.
 # Manifests under test: every PrometheusRule in apps/wave-0/ — gitea, jackett, pihole, plex,
-# qbittorrent (container/Deployment) and vm-demo (KubeVirt VMI). The harness discovers this file by
+# qbittorrent (container/Deployment), pihole (container/StatefulSet — `availability.replicas: 2`, so
+# its SLI reads kube_statefulset_status_replicas_ready; see schemas/renderer-contract.md
+# `### availability`) and vm-demo (KubeVirt VMI). The harness discovers this file by
 # the SUITE convention (`<dir>/tests/*.test.yaml`) and flattens those RENDERED manifests, so these
 # tests run against the exact expressions Flux applies, never a copy.
 #
@@ -41,7 +43,7 @@ tests:
        values: '1+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="jackett", deployment="jackett"}'
        values: '1+0x20'
-      - series: 'kube_deployment_status_replicas_available{namespace="pihole", deployment="pihole"}'
+      - series: 'kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"}'
        values: '1+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="plex", deployment="plex"}'
        values: '1+0x20'
@@ -81,7 +83,7 @@ tests:
        values: '1+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="jackett", deployment="jackett"}'
        values: '1+0x20'
-      - series: 'kube_deployment_status_replicas_available{namespace="pihole", deployment="pihole"}'
+      - series: 'kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"}'
        values: '1+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="plex", deployment="plex"}'
        values: '0+0x20'
@@ -132,7 +134,7 @@ tests:
        values: '0+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="jackett", deployment="jackett"}'
        values: '0+0x20'
-      - series: 'kube_deployment_status_replicas_available{namespace="pihole", deployment="pihole"}'
+      - series: 'kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"}'
        values: '0+0x20'
      - series: 'kube_deployment_status_replicas_available{namespace="plex", deployment="plex"}'
        values: '1+0x20'
@@ -170,10 +172,10 @@ tests:
              severity: critical
              service: pihole
              namespace: pihole
-              deployment: pihole
+              statefulset: pihole
            exp_annotations:
              summary: "pihole is down"
-              description: "pihole: its Deployment has no available/ready replica for 5m, or the signal is gone entirely (SLI: up, via kube_deployment_status_replicas_available)."
+              description: "pihole: its StatefulSet has no available/ready replica for 5m, or the signal is gone entirely (SLI: up, via kube_statefulset_status_replicas_ready)."
      - eval_time: 10m
        alertname: qbittorrent-down
        exp_alerts:
@@ -153,11 +153,16 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://radarr.${BASE_DOMAIN}/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --api-route=^/api/
+        - --api-route=^/signalr/
+        - --cookie-name=_oauth2_proxy_radarr
+        - --cookie-domain=radarr.${BASE_DOMAIN}
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -178,6 +183,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -275,6 +283,28 @@ spec:
      - port: '9117'
        protocol: TCP
 ---
+# CiliumNetworkPolicy egress allow (R-NET-02): radarr-allow-plex-32400
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: radarr-allow-plex-32400
+  namespace: radarr
+  labels:
+    app.kubernetes.io/name: radarr
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: radarr
+  egress:
+  - toEndpoints:
+    - matchLabels:
+        app.kubernetes.io/name: plex
+        k8s:io.kubernetes.pod.namespace: plex
+    toPorts:
+    - ports:
+      - port: '32400'
+        protocol: TCP
+---
 # CiliumNetworkPolicy reciprocal ingress allow (R-NET-02): radarr-allow-seerr-7878
 apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
@@ -153,11 +153,14 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://sonarr.${BASE_DOMAIN}/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_sonarr
+        - --cookie-domain=sonarr.${BASE_DOMAIN}
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -178,6 +181,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -275,6 +281,28 @@ spec:
      - port: '9117'
        protocol: TCP
 ---
+# CiliumNetworkPolicy egress allow (R-NET-02): sonarr-allow-plex-32400
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: sonarr-allow-plex-32400
+  namespace: sonarr
+  labels:
+    app.kubernetes.io/name: sonarr
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: sonarr
+  egress:
+  - toEndpoints:
+    - matchLabels:
+        app.kubernetes.io/name: plex
+        k8s:io.kubernetes.pod.namespace: plex
+    toPorts:
+    - ports:
+      - port: '32400'
+        protocol: TCP
+---
 # CiliumNetworkPolicy reciprocal ingress allow (R-NET-02): sonarr-allow-seerr-8989
 apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
@@ -51,6 +51,23 @@ auth:
    # This grants no real privilege: platform-admins already maps to cluster-admin via
    # k8s-oidc-rbac.yaml (oidc:platform-admins), so a member can already exec into these pods.
    - platform-admins           # operators of these services must not be locked out of them
+  # Answer Jackett's OWN API with 401 instead of a 302 to Keycloak (schema 1.9.0; renders
+  # --api-route, singular and repeatable, onto the oauth2-proxy sidecar).
+  #
+  # MEASURED on this pod's access log, 2026-09-30: on /api/ paths, 33 x 302 against 2 x 401 —
+  # and 4 of 6 POSTs were answered BY THE PROXY and never reached Jackett (access-log upstream
+  # field `-`, not `/`), including both `/api/v2.0/server/update` calls. Jackett's jQuery does
+  # not send the exact `Accept: application/json` oauth2-proxy's AJAX guard requires, so the
+  # 302 was followed invisibly, satisfied silently by the live Keycloak SSO session, and
+  # resolved 200 carrying the dashboard's HTML where the JavaScript expected JSON. A browser
+  # following a 302 also re-issues a POST as a GET and DROPS THE BODY, which is why "adding an
+  # indexer silently does nothing" was the user-visible symptom and nothing logged a failure.
+  #
+  # SCOPE: `/api/` only. Jackett serves its UI at /UI/ and its static assets elsewhere, and a
+  # browser never NAVIGATES to /api/ — so a plain page load is still 302'd into a silent SSO
+  # sign-in, which is the behaviour that must not change.
+  apiRoutes:
+    - ^/api/
 update:
  channel: stable               # R-UPD-01
  range: ">=0.22.0"             # stay on the Jackett 0.22.x stable line
@@ -12,6 +12,42 @@ storage:
  - name: config
    size: 2Gi
    class: longhorn           # replicated config state; DR-critical (docs/04 line 118; R-STATE-02/03)
+    # THE PVC WAS MOUNTED SOMEWHERE PIHOLE NEVER WRITES, AND SO HELD NOTHING AT ALL. Without an
+    # explicit mountPath the renderer defaults to /data/<name>, i.e. /data/config — but PiHole keeps
+    # every piece of its state in /etc/pihole. MEASURED on live prod 2026-10-04: `/etc/pihole` (7.1 M
+    # — gravity.db, pihole.toml, the query DB) was on the **overlay** filesystem, i.e. the
+    # container's EPHEMERAL layer, while `/dev/longhorn/pvc-849edb86...` sat mounted at
+    # /data/config holding `lost+found` and 24 K of nothing. Three consequences, all live until now:
+    #   * every restart silently discarded PiHole's whole state, re-downloading 72 k gravity domains
+    #     and regenerating the admin password hash;
+    #   * `backup.policy: daily-offcluster` below has been faithfully backing up an EMPTY volume,
+    #     which is worse than no backup because it reports success;
+    #   * it is also why phase 1's password/adlist declaration mattered more than it first appeared —
+    #     that state was not merely undeclared, it was ephemeral.
+    # This is the same class of bug catalog/plex.yaml's `config` entry documents ("Without this the
+    # PVC sat unused and Plex wrote to the container's EPHEMERAL layer").
+    #
+    # /etc/pihole IS THE IMAGE'S INTENDED MOUNT POINT, and an EMPTY volume there self-initialises.
+    # Verified in the running image, not assumed:
+    #   * `/usr/bin/start.sh` (PID 1) calls `migrate_gravity`, whose `migrate_to_database()` creates
+    #     the gravity database when `[ ! -e "$gravityDBfile" ]` — the same branch the postStart hook
+    #     below relies on, and the branch that produced the current ephemeral copy.
+    #   * FTL writes `pihole.toml` itself from the `FTLCONF_*` env above (that is what the
+    #     `### CHANGED (env)` markers in it are).
+    #   * NO `podSecurityContext.fsGroup` is needed and adding one would be inert: `start.sh` line 5
+    #     is `chown -R pihole:pihole /etc/pihole/ /var/log/pihole/` and PID 1 runs as **uid 0**, so
+    #     the image fixes ownership of a fresh root-owned volume before anything else starts. Same
+    #     arrangement as plex's image managing its own /config ownership.
+    #   * the image's own default for `files.macvendor` was moved OUT of this directory (to
+    #     `/macvendor.db`), which is what a maintainer does to a path they expect to be a volume.
+    #
+    # WHY IT IS FIXED IN THIS BRANCH rather than left for its own: `availability.replicas: 2` renders
+    # per-replica volumeClaimTemplates, and with the wrong mountPath those would be two 2 Gi Longhorn
+    # volumes (times three Longhorn replicas each) holding nothing but `lost+found` while all state
+    # stayed ephemeral — paying the cost of persistence and getting none of it. It also REMOVES this
+    # change's main operational risk: with gravity.db persisted, a pod restart is cheap, instead of
+    # every restart re-downloading 72 k domains inside a wave that has a 5 m timeout.
+    mountPath: /etc/pihole
 dependencies: []              # default-deny: reaches only DNS; it PROVIDES DNS (docs/03; R-NET-01/02)
 # ------------------------------------------------------------------------------------------------
 # LAN SPLIT-HORIZON DNS — the half of R-DNS-03 that platform/coredns/coredns-custom.yaml defers to.
@@ -100,6 +136,53 @@ env:
  # while resolution for the zone does not work. Confirmed both ways against the live image.
  - name: FTLCONF_misc_dnsmasq_lines
    value: "address=/${BASE_DOMAIN}/${GATEWAY_VIP};address=/unifi.${BASE_DOMAIN}/${UNIFI_L4_VIP};address=/crafty-mc.${BASE_DOMAIN}/${CRAFTY_L4_VIP};address=/pihole-dns.${BASE_DOMAIN}/${PIHOLE_DNS_VIP}"
+  # THE ADMIN PASSWORD, previously the ONE operator-set value declared nowhere. Sourced from
+  # apps/wave-0/pihole-bootstrap-secret.sops.yaml (SOPS/age, same wave => same reconcile), which
+  # carries the full rationale and the measurement that found it.
+  #
+  # `webserver.api.password` takes CLEARTEXT and FTL hashes it itself into `webserver.api.pwhash`.
+  # Verified in the running image, not assumed: `pihole setpassword` is itself
+  # `setFTLConfigValue "webserver.api.password" "$PASSWORD"`, commented upstream as "pihole-FTL will
+  # automatically hash the password". So the BALLOON-SHA256 hash on the PVC is now a DERIVED artefact
+  # and this is its source of truth. Rotating the Secret rotates the password.
+  - name: FTLCONF_webserver_api_password
+    valueFrom:
+      secretKeyRef:
+        name: pihole-bootstrap
+        key: webui-password
+availability:
+  # TWO REPLICAS, ON DIFFERENT NODES. LAN DNS is a critical network service and it was ONE pod on
+  # ONE node. Measured 2026-10-04, that cost more than "until it reschedules":
+  #   * the node running the pod dies -> ~40s to mark it NotReady, then `tolerationSeconds: 300`
+  #     before the pod is even evicted, then reschedule, then Longhorn must force-detach the RWO
+  #     volume from a dead node and reattach it. Roughly 6-10 MINUTES with no LAN resolution.
+  #   * the node holding the `pihole-dns-*` L2 lease dies -> Cilium re-elects an announcer, tens of
+  #     seconds. (These were DIFFERENT nodes: lease on k3s-prod-2, pod on k3s-prod-1. Two of three
+  #     nodes were single points of failure for LAN DNS, with wildly different blast radii.)
+  # The Services already tolerate this: `externalTrafficPolicy: Cluster` on the DNS LoadBalancers
+  # (apps/wave-0/pihole-dns.yaml) means the L2 lease holder forwards to ANY ready backend, so a
+  # second replica is picked up with no change there.
+  #
+  # WHAT THIS DOES NOT FIX, and it is the bigger risk: LAN DNS still depends on the CLUSTER. etcd
+  # quorum loss, a Cilium regression, a Longhorn problem or a bad reconcile takes DNS down
+  # everywhere regardless of replica count. The only thing that removes that dependency is a
+  # resolver outside the cluster carrying the same `address=` lines. Considered and not chosen
+  # (operator decision, 2026-10-04); see decisions/2026-10-04-pihole-multi-replica.md.
+  #
+  # THE PREREQUISITE WAS DONE FIRST, and this field would have been actively harmful without it. A
+  # replica gets a FRESH volume, and until 2026-10-04 PiHole's admin password and adlists lived only
+  # on the PVC — so replica 1 would have come up SERVING DNS AND BLOCKING NOTHING, with a different
+  # password, and answers would have varied by which pod replied. Both are now declared (the
+  # password as a SOPS Secret -> FTLCONF_webserver_api_password; the adlists reconciled into
+  # gravity.db by the postStart hook below), which is what makes a fresh replica identical to its
+  # sibling. See decisions/2026-10-04-pihole-declare-state.md.
+  #
+  # WHY 2 AND NOT 3. Two tolerates the loss of any single node, which is the stated goal, and leaves
+  # a spare node for the evicted replica to land on under the HARD anti-affinity the renderer emits.
+  # Three would pin one replica per node with no spare, so a single node failure would leave a
+  # replica permanently Pending and — `apps-wave-0` being `wait: true` — stall the app tier for as
+  # long as the node was down. Two is the count that degrades gracefully here.
+  replicas: 2
 ingressRoutes:
  - host: pihole.${BASE_DOMAIN}
    port: 80                  # admin web UI, protected at the Gateway by forward-auth (docs/09)
@@ -107,6 +190,117 @@ auth:
  mode: forward-auth          # oauth2-proxy -> Keycloak; bare UI, no native OIDC (docs/15; R-IDP-02)
  groups:
    - platform-admins         # sensitive admin UI: authenticated AND in group (R-IDP-06)
+  # NO `apiRoutes` HERE, DELIBERATELY. PiHole's admin UI serves no /api/ path, so the `^/api/`
+  # that jackett and radarr need would match nothing on a good day — and a BROADER guess would
+  # answer PiHole's own page loads with 401, leaving a logged-out operator on a bare error page
+  # with no sign-in offered. This app is the reason auth.apiRoutes is per-service (schema 1.9.0)
+  # rather than a renderer constant: the measurement that justifies the flag on two apps is
+  # evidence about those two apps. Its oauth2-proxy log holds 0 unauthenticated requests across
+  # the current pod's whole 23h lifetime (measured 2026-10-01), so there is nothing to act on.
+lifecycle:
+  # LEVEL-TRIGGERED ADLIST ASSERTION, run on EVERY container start. Same pattern, and the same
+  # reasoning, as catalog/qbittorrent.yaml's postStart: desired state lives in git and is
+  # RE-ASSERTED unconditionally rather than set once by hand.
+  #
+  # WHY THIS IS NEEDED AT ALL — the blocklists were NOT reproducible from git. Measured on live prod
+  # 2026-10-04: `gravity.db` held exactly ONE adlist (StevenBlack's hosts, enabled) and 72,234
+  # gravity domains, and NOTHING in this repo named it. Worse, a FRESH database does not seed it:
+  # the image's own schema template `/etc/.pihole/advanced/Templates/gravity.db.sql` creates the
+  # `adlist` table and inserts NO rows into it (grepped in the running image — there is no
+  # `INSERT INTO adlist` at all). So a rebuild from git, or a replica with a new PVC, would come up
+  # SERVING DNS AND BLOCKING NOTHING, silently and with every object reporting healthy.
+  #
+  # WHY A postStart HOOK AND NOT A FILE. PiHole v6 keeps adlists as DB rows, not config. The
+  # `/etc/pihole/adlists.list` file still exists in `gravity.sh` but is MIGRATION-ONLY: its import
+  # sits inside `migrate_to_database()`, guarded by `if [ ! -e "$gravityDBfile" ]`, so it fires only
+  # while CREATING a database and never again on an existing one. A ConfigMap mounted there would
+  # therefore be silently ignored on every pod this cluster currently runs. There is also no
+  # `pihole adlist` CLI subcommand (only allow/deny for domains), so SQL against the DB is the
+  # actual interface. FTL ships an embedded shell (`pihole-FTL sqlite3`, SQLite 3.53.1) so no extra
+  # tooling is needed in the image.
+  #
+  # IT RECONCILES, IT DOES NOT JUST ADD: adlists present in the DB but absent from DECLARED below are
+  # DISABLED (not deleted — deleting cascades to `gravity` rows and would force a full re-download of
+  # everything on a typo). So this list is the source of truth in both directions, and `pihole -g`
+  # runs ONLY when something actually changed, which keeps the steady-state start free of a 70k-domain
+  # rebuild.
+  #
+  # NO BRACED DOLLAR EXPANSION ANYWHERE IN THIS SCRIPT, DELIBERATELY — the same contract
+  # catalog/qbittorrent.yaml documents at length. apps-wave-0 declares postBuild.substituteFrom
+  # cluster-vars, so Flux runs envsubst over this whole manifest at reconcile time; a `${VAR}` form
+  # would behave differently on the cluster than under a plain `kubectl apply` of the golden. Bare
+  # `$VAR` and `$(cmd)` are untouched by envsubst and so behave identically however this is run.
+  # `render.py`'s assert_substitution_safe refuses to emit a manifest that breaks this.
+  postStart:
+    exec:
+      command:
+        - /bin/bash
+        - -c
+        - |
+          # Assert the declared adlists in gravity.db. Idempotent: steady state writes nothing and
+          # does NOT run a gravity rebuild.
+          set -u
+          GRAVITY=/etc/pihole/gravity.db
+
+          # THE DECLARED SET — one URL per line, the source of truth for what PiHole blocks.
+          # To add a list: add a line here, commit, let Flux reconcile. Nothing else to do.
+          DECLARED='https://raw.githubusercontent.com/StevenBlack/hosts/master/hosts'
+
+          # On a fresh PVC, FTL creates gravity.db a moment after the container starts. Wait, bounded,
+          # and NEVER block or fail the start: a missing DB here must not crashloop DNS for the LAN.
+          i=0
+          while [ ! -s "$GRAVITY" ] && [ "$i" -lt 90 ]; do sleep 2; i=$(expr $i + 1); done
+          if [ ! -s "$GRAVITY" ]; then
+            echo "pihole-adlists: gravity.db absent after 180s - leaving adlists untouched" >&2
+            exit 0
+          fi
+
+          sql() { pihole-FTL sqlite3 "$GRAVITY" "$1"; }
+          CHANGED=0
+
+          # 1) Every declared URL must be present AND enabled.
+          echo "$DECLARED" | while IFS= read -r url; do
+            [ -n "$url" ] || continue
+            printf '%s\n' "$url"
+          done > /tmp/declared.txt
+          while IFS= read -r url; do
+            [ -n "$url" ] || continue
+            # Single quotes inside a URL would break this SQL; adlist URLs do not contain them.
+            have=$(sql "select count(*) from adlist where address='$url';")
+            if [ "$have" = "0" ]; then
+              echo "pihole-adlists: adding $url"
+              sql "insert into adlist (address,enabled,comment) values ('$url',1,'declared in catalog/pihole.yaml');"
+              CHANGED=1
+            else
+              en=$(sql "select enabled from adlist where address='$url';")
+              if [ "$en" != "1" ]; then
+                echo "pihole-adlists: re-enabling $url"
+                sql "update adlist set enabled=1 where address='$url';"
+                CHANGED=1
+              fi
+            fi
+          done < /tmp/declared.txt
+
+          # 2) Anything enabled in the DB but NOT declared gets disabled (never deleted - see above).
+          sql "select address from adlist where enabled=1;" | while IFS= read -r addr; do
+            [ -n "$addr" ] || continue
+            if ! grep -qxF "$addr" /tmp/declared.txt; then
+              echo "pihole-adlists: disabling undeclared $addr"
+              sql "update adlist set enabled=0 where address='$addr';"
+              echo changed >> /tmp/adlist-changed
+            fi
+          done
+          [ -s /tmp/adlist-changed ] && CHANGED=1
+          rm -f /tmp/adlist-changed /tmp/declared.txt
+
+          # 3) Rebuild gravity ONLY if the declared set actually moved.
+          if [ "$CHANGED" = "1" ]; then
+            echo "pihole-adlists: declared set changed - running gravity update"
+            pihole -g || echo "pihole-adlists: gravity update failed; existing gravity.db retained" >&2
+          else
+            echo "pihole-adlists: already matches the declared set - nothing to do"
+          fi
+          exit 0
 update:
  channel: stable             # R-UPD-01
  range: ">=2024.0.0"
@@ -72,14 +72,108 @@ env:
      secretKeyRef:
        name: plex-claim
        key: PLEX_CLAIM
-  # ADVERTISE_IP ("Custom server access URLs") is DEFERRED, not accidentally omitted. It tells Plex
-  # which URL to advertise to plex.tv for REMOTE clients. The only candidate is
-  # https://plex.home.keithjackson.dev, but that host resolves via split-horizon DNS to the INTERNAL
-  # Gateway VIP (conformance/tests/l2-behavioral/dns-internal-resolution), and both its public
-  # reachability and whether the TLS-terminating Cilium Gateway proxies Plex's 32400 API correctly for
-  # remote access are UNVERIFIED. Advertising a wrong/internal-only URL breaks remote access for ALL
-  # clients, so per "prefer safety over completeness" it is left unset. FOLLOW-UP: once remote
-  # reachability through the Gateway is confirmed, set ADVERTISE_IP: "https://plex.home.keithjackson.dev:443".
+  # ADVERTISE_IP ("Custom server access URLs") — NOW SET. This replaces a DEFERRAL note that stood
+  # here until 2026-10-04, and the deferral's premise has changed: it weighed only ONE candidate,
+  # `https://plex.home.keithjackson.dev` through the L7 Gateway. The operator chose the NATIVE L4
+  # path instead on 2026-10-03 (apps/wave-0/plex-remote.yaml) so streams go straight to
+  # ${PLEX_L4_VIP}:32400 and bypass the Envoy hop. What was still missing was a URL whose TLS
+  # CERTIFICATE validates — and getting that wrong is what took the service down.
+  #
+  # THE VALUE IS A `plex.direct` NAME, NOT THE RAW VIP, AND THAT IS THE WHOLE POINT. Plex serves
+  # 32400 with its own certificate; measured live 2026-10-04 it carries exactly one name:
+  #     CN = *.<serverhash>.plex.direct      (one SAN, no IP SANs at all)
+  # so `https://<ip>:32400` can NEVER validate. `<dashed-ip>.<serverhash>.plex.direct` is Plex's own
+  # mechanism for this: it is PUBLIC DNS that resolves to the embedded address, so the name points at
+  # the LAN VIP while the wildcard cert above matches it. Measured 2026-10-04 with FULL verification
+  # (no `curl -k`):
+  #     https://192-168-2-24.<hash>.plex.direct:32400/identity -> 200, TLS verify OK
+  #     https://192.168.2.24:32400/identity                    -> 000, cert validation FAILED
+  #
+  # THIS FIELD EXISTS BECAUSE AUTO-DISCOVERY CANNOT WORK HERE. Left empty, Plex publishes the
+  # addresses it can see itself — which from a pod is the overlay IP (`10.0.2.136`), useless to a LAN
+  # client. The VIP has to be advertised explicitly; there is no "just clear it" fix.
+  #
+  # WHAT WENT WRONG, recorded so it is not re-introduced by hand. customConnections OVERRIDES Plex's
+  # own publishing, and on 2026-10-04 it held two hand-typed URLs, both uncompletable:
+  #   - `https://192.168.2.24:32400`              -> cert cannot cover an IP. Measured: a storm of
+  #     `CERT: incomplete TLS handshake ... sslv3 alert certificate unknown` in Plex's log every few
+  #     seconds, from REAL LAN clients arriving through the VIP.
+  #   - `https://plex.home.keithjackson.dev:32400` -> nothing listens. The Gateway serves this host on
+  #     443; measured `http=000` against 192.168.2.20:32400.
+  # With every advertised connection dead, plex.tv had nothing publishable, sat at
+  # `mapping state 'Mapped - Not Published'`, fell back to Plex Relay, and THE SERVER DISAPPEARED
+  # FROM THE ACCOUNT — clients reported "not authorized".
+  #
+  # DECLARING IT HERE rather than in the web UI is the durable half of the fix. The image's
+  # /etc/cont-init.d/40-plex-first-run runs `setPref "customConnections" "${ADVERTISE_IP}"` (line 127)
+  # OUTSIDE the PLEX_CLAIM/PlexOnlineToken guard, and its `/.firstRunComplete` marker sits on the
+  # container's OVERLAY fs, not the /config PVC — so this is re-applied on EVERY pod start and git,
+  # not a form field, owns the value. Editing this env rolls the Deployment, which is what applies it.
+  #
+  # TWO LIMITS, stated rather than discovered later:
+  #   - THE HASH IS SERVER IDENTITY, which is why the whole URL is a cluster-var and not built from
+  #     ${PLEX_L4_VIP}: envsubst cannot dash an IP, and the hash is not derivable from anything in
+  #     this tree. It is not a secret (it ships in the cert and in the published connection list), but
+  #     if Plex reissues under a new hash this URL goes stale and must be re-read from the live cert.
+  #   - THIS RESTORES LAN CLIENTS AND THE ACCOUNT LISTING, NOT WAN. Remote clients still need the
+  #     off-repo tcp/32400 port-forward (runbooks/plex-remote-access.md). Once that exists, the WAN
+  #     twin of this URL can be appended comma-separated — customConnections takes a list.
+  - name: ADVERTISE_IP
+    value: ${PLEX_ADVERTISE_URL}
+# HARDWARE TRANSCODING — the host iGPU's render device (schema 1.8.0 `hardware`; R-PLAT-02).
+# INTENT ONLY, AND THE MECHANISM DELIBERATELY DOES NOT LIVE HERE. This says "this workload needs an
+# Intel Quick Sync render device"; it does NOT name a device path and it does NOT name a resource.
+# How the platform hands the device over is `catalog/render.py`'s business and is documented in
+# `schemas/renderer-contract.md` `### hardware` — read that for the mechanism, never this comment.
+# The payoff is proven, not theoretical: on 2026-10-01 the mechanism was swapped outright (hostPath
+# -> Intel GPU device plugin extended resource) and THIS BLOCK DID NOT MOVE — a `render.py` change
+# plus a re-render, zero catalog edits. Do not "simplify" it into a raw `/dev/...` path; that is the
+# whole point of the enum.
+#
+# THE DEVICE IS PRESENT ON EVERY PROD NODE, so nothing had to be installed and Plex can land
+# anywhere. Measured on all three nodes 2026-09-30: `/dev/dri/renderD128` present, `root:render`,
+# GID 992, mode 0660, driver `i915` bound (NOT `xe`), UHD Graphics 630 on an i7-9700T, media engines
+# `vcs0`/`vecs0` live, iGPU enabled in BIOS on each. Measured again 2026-10-01, one layer up: all
+# three nodes advertise `gpu.intel.com/i915: 1` in `.status.allocatable`. This field renders no
+# nodeSelector; placement is a separate concern with its own justification.
+#
+# PLEX'S SCHEDULABILITY IS NOW COUPLED TO THE `intel-gpu-plugin` DaemonSet, and that is a real cost,
+# not a footnote. The current mechanism is a hard scheduling constraint: if the resource is
+# unavailable (plugin broken, removed, or not yet reconciled) this pod is UNSCHEDULABLE, and because
+# `apps-wave-0` is `wait: true` with `apps-wave-1` `dependsOn` it, that STALLS THE WHOLE APP TIER.
+# The previous hostPath mechanism degraded silently to software transcoding instead — but it never
+# actually granted access (device-cgroup EPERM, measured 2026-10-01), so there was nothing to keep.
+# THE ESCAPE HATCH IS THIS BLOCK: delete these two lines and re-render, and Plex goes back to
+# software transcoding and the tier unblocks. One field, no code change. `runbooks/plex-quicksync.md`
+# carries that procedure for whoever needs it at 2am.
+#
+# NO `podSecurityContext.supplementalGroups`, AND ADDING ONE WOULD BE INERT — do not "fix" this.
+# The device is `root:render` 0660 and this process is uid/gid 1000, so the obvious move is to push
+# the host `render` GID in as a supplemental group. It cannot work here, and this survived the
+# mechanism swap unchanged. MEASURED IN THE RUNNING CONTAINER, 2026-09-30 and again 2026-10-01: the
+# image's own `/etc/cont-init.d/45-plex-hw-transcode-and-connected-tuner` stats every char device
+# under `/dev/dri`, `groupadd`s a matching group INSIDE the container when none exists, and
+# `usermod -a -G`s the `plex` user into it — every start, nothing hardcoded, which also makes a host
+# GID renumbering self-healing (no GID is pinned anywhere in this repo) and which works the same way
+# on a PLUGIN-INJECTED device as on a bind-mounted one. Measured live, `Plex Media Server` ran with
+# `Groups: 44 100 992 1000` — 992 IS the host `render` GID, i.e. the hook had already done the job.
+# The service execs via `s6-setuidgid plex`, and `s6-setuidgid` REPLACES the process group list from
+# `/etc/group` rather than appending to it, so root's gid `0` — which PID 1 (`s6-svscan`) does hold —
+# is GONE, and anything Kubernetes supplies reaches s6-svscan and is discarded before Plex ever
+# starts. The group is the image's job and it does it correctly. `render.py --selftest` asserts the
+# ABSENCE of a pod securityContext for this field precisely so nobody adds an inert one back later.
+#
+# THE LAST MILE IS A MANUAL OPERATOR STEP AND CANNOT LIVE IN THIS REPO. Handing over the device does
+# not enable hardware transcoding — Plex's own `HardwareAcceleratedCodecs` preference does, and it
+# is ABSENT from `Preferences.xml` today (the toggle has never been on). NO environment variable can
+# set it: a grep of the running image's entire init surface (`/etc/cont-init.d`, `/etc/services.d`,
+# `/etc/fix-attrs.d`, `/plex-common.sh`, `/usr/local/bin`) found ZERO `PLEX_PREFERENCE*` support,
+# and the six prefs the image does write (`MachineIdentifier`, `ProcessedMachineIdentifier`,
+# `PlexOnlineToken`, `customConnections`, `allowedNetworks`, `TranscoderTempDirectory`) are all
+# unrelated. So there is no GitOps path to the toggle; an operator clicks it once in the web UI.
+# `runbooks/plex-quicksync.md` is that procedure, plus how to verify it took.
+hardware:
+  gpu: intel-quicksync
 exposure: public              # edge port-forward + public DNS; clients + Seerr reach it (docs/09; R-DNS/R-NET)
 egress: internet              # normal WAN egress (docs/03; R-NET-05)
 storage:
@@ -448,6 +448,12 @@ auth:
    # via k8s-oidc-rbac.yaml). Without it, an empty media-users locks every operator out with an
    # immediate 403 at the oauth2-proxy callback — observed on prod 2026-09-12.
    - platform-admins           # operators of these services must not be locked out of them
+  # NO `apiRoutes` HERE: UNMEASURED, not cleared. qBittorrent's WebUI drives everything through
+  # /api/v2/, so it is plausibly affected in the same way jackett is — but plausibility is what
+  # put a renderer constant on the table in the first place. MEASURED 2026-10-01 on the live pod
+  # (qbittorrent-5547bdb9bd-zq557, 0 restarts, 23h): 13,649 log lines, ZERO unauthenticated
+  # requests, nothing but health probes. No traffic means no evidence either way. Add `^/api/`
+  # when someone opens the WebUI and records the 302/401 split, as was done for jackett.
 update:
  channel: stable               # R-UPD-01
  range: ">=5.0.0 <6.0.0"       # stay on the qBittorrent 5.x stable line
@@ -100,6 +100,9 @@ dependencies:
  - to: jackett                 # arr → indexer proxy (docs/13 dependency graph; R-NET-02 explicit allow)
    port: 9117
    protocol: tcp
+  - to: plex
+    port: 32400                 # arr -> plex: required since library is on NFS which does not provide file CRUD notifications
+    protocol: tcp
 ingressRoutes:
  - host: radarr.${BASE_DOMAIN}
    port: 7878                  # Radarr web/API; TLS terminates at the shared Gateway (docs/08, docs/09)
@@ -112,6 +115,32 @@ auth:
    # via k8s-oidc-rbac.yaml). Without it, an empty media-users locks every operator out with an
    # immediate 403 at the oauth2-proxy callback — observed on prod 2026-09-12.
    - platform-admins           # operators of these services must not be locked out of them
+  # Answer Radarr's machine endpoints with 401 instead of a 302 to Keycloak (schema 1.9.0;
+  # renders --api-route, singular and repeatable, onto the oauth2-proxy sidecar).
+  #
+  # `^/signalr/` — MEASURED ON THE LIVE POD, 2026-10-01, 23h after the cookie fix deployed
+  # (radarr-7d55cbffb9-42hmh, 0 restarts): 232 unauthenticated requests in the pod's whole
+  # lifetime and ALL 232 are the same call — `POST /signalr/messages/negotiate` from a browser
+  # UA, answered 302 by the proxy with upstream `-`, i.e. Radarr never saw one of them. That is
+  # a left-open tab whose session the cookie rename orphaned, polling SignalR forever, silently
+  # redirected every time and NEVER re-prompting the operator. It is the reported symptom
+  # ("reloading never fixes it and I am never re-prompted") still running in production on a
+  # service nobody was watching. With 401 the UI's connection breaks LOUDLY, the operator
+  # reloads, and `/` is outside both regexes so the reload signs them back in.
+  #
+  # `^/api/` — measured on radarr, but on the PRE-FIX pod and NOT by me: the 2026-09-30
+  # diagnosis (decisions/2026-09-30-jackett-oauth2-proxy-session.md) recorded radarr's 156
+  # unauthenticated requests as `/api/v3/command` (62), `/api/v3/queue/details` (46) and
+  # `/signalr/messages/negotiate` (43). The current pod has logged ZERO /api/ traffic, so I did
+  # not re-measure it; it is included because 108 of 156 measured rejections on this app were
+  # /api/v3 XHRs, which is this app's own evidence rather than an analogy to another app.
+  #
+  # SCOPE: both are JSON/RPC namespaces a browser never NAVIGATES to. Radarr's UI is served at
+  # `/` with assets under /Content/ and /MediaCover/, none of which match — so a plain page load
+  # is still 302'd into a silent SSO sign-in.
+  apiRoutes:
+    - ^/api/
+    - ^/signalr/
 update:
  channel: stable               # R-UPD-01
  range: ">=5.0.0 <6.0.0"       # stay on the Radarr v5 stable line
@@ -100,6 +100,9 @@ dependencies:
  - to: jackett                 # arr → indexer proxy (docs/13 dependency graph; R-NET-02 explicit allow)
    port: 9117
    protocol: tcp
+  - to: plex                    # arr -> plex: required since library is on NFS which does not provide file CRUD notifications
+    port: 32400
+    protocol: tcp
 ingressRoutes:
  - host: sonarr.${BASE_DOMAIN}
    port: 8989                  # Sonarr web/API; TLS terminates at the shared Gateway (docs/08, docs/09)
@@ -112,6 +115,14 @@ auth:
    # via k8s-oidc-rbac.yaml). Without it, an empty media-users locks every operator out with an
    # immediate 403 at the oauth2-proxy callback — observed on prod 2026-09-12.
    - platform-admins           # operators of these services must not be locked out of them
+  # NO `apiRoutes` HERE, AND NOT BECAUSE IT NEEDS NONE — because nothing measures that it does.
+  # Sonarr is nearly the same codebase as Radarr, which is exactly what makes `^/api/` +
+  # `^/signalr/` a tempting guess and still a guess. MEASURED 2026-10-01 on the live pod
+  # (sonarr-ddf485775-8ckwt, 0 restarts, 23h): its oauth2-proxy log contains 13,643 lines and
+  # ZERO unauthenticated requests — no /api/, no /signalr/, nothing but health probes. Nobody
+  # has opened this UI since the cookie fix rolled, so the measurement is empty rather than
+  # negative: it is evidence of no TRAFFIC, not evidence of no need. Add the routes when someone
+  # reproduces the 302-on-XHR here, with the numbers.
 update:
  channel: stable               # R-UPD-01
  range: ">=4.0.0 <5.0.0"       # stay on the Sonarr v4 stable line
@@ -0,0 +1,194 @@
+#!/usr/bin/env python3
+"""ci/approval-override.py — select the admin-override on a PR, if there is a valid one.
+
+This script decides NOTHING about authority. It parses a PR's comments and reviews and returns
+the single override *candidate* that ci/approval-policy.sh should then try to authorise against
+the live Gitea permission API. Keeping the two apart is the point:
+
+  * parsing is pure, offline, and therefore TESTABLE (ci/tests/approval-override.test.sh);
+  * authorisation needs the API and must stay in the stage that fails closed.
+
+A candidate is a comment OR a review whose body contains, AT THE START OF A LINE, the directive
+followed by a non-empty reason:
+
+    /approve-override the gate cannot gate itself and I am the only human reviewer
+
+TWO SOURCES, AND WHY THAT IS NOT BELT-AND-BRACES. Measured 2026-10-02 on PR #38: the override was
+posted as an ordinary PR comment, was well-formed, was newer than the head commit — and the stage
+still blocked, because CI could not READ it. `SETTINGS_API_TOKEN` is minted with scopes
+`[read:repository, read:admin]`, and Gitea puts /issues/{n}/comments behind the *issue* scope, so
+the request came back as an error OBJECT rather than an array. `/pulls/{n}/reviews` is under
+read:repository and was readable in the same run. So:
+
+  * a REVIEW body works with the token CI has today (leave a review of type "Comment");
+  * a COMMENT body works once the token carries read:issue (see infra/gitea/set-ci-token.sh).
+
+Accepting both means the hatch is not hostage to one API scope. A source that cannot be read
+contributes no candidates and is reported on stderr; it can only cause a false BLOCK, never a
+false pass, which is why degrading it is safe and blocking on it was not necessary.
+
+WHY EACH RULE IS HERE
+
+  * START OF A LINE, not "anywhere in the body". Otherwise quoting someone else's override
+    ("you could just /approve-override this") re-arms it. Lines beginning with '>' are DROPPED
+    rather than unquoted, for exactly that reason.
+  * A REASON IS MANDATORY. A bare directive records that a review was waived but not why, which
+    is the difference between an audit trail and a rubber stamp.
+  * STALE OVERRIDES ARE DISCARDED. An override earned at commit A must not silently cover commit
+    B. Any candidate created at or before the head commit's timestamp is dropped — the same
+    semantics Gitea's own `dismiss_stale_approvals` applies to approvals. Push new commits and
+    you must re-override.
+  * A PENDING REVIEW IS NOT SUBMITTED. It is a private draft visible only to its author, so it is
+    not an act of anything and is dropped.
+  * THE LATEST CANDIDATE WINS, across both sources, so a later override supersedes an earlier one.
+
+Output: one TSV line (login, timestamp, source, reason) on stdout, or nothing at all when there is
+no valid candidate. Never exits non-zero for "no override" — absence is a normal answer and the
+caller must not be able to confuse it with a parse crash.
+
+Input, all via env so nothing lands in `ps`:
+  COMMENTS_JSON  the Gitea /issues/{n}/comments array (or an error object; or absent)
+  REVIEWS_JSON   the Gitea /pulls/{n}/reviews array   (or an error object; or absent)
+  HEAD_TIME      ISO-8601 timestamp of the PR head commit
+  DIRECTIVE      the override directive, e.g. /approve-override
+"""
+import json
+import os
+import re
+import sys
+from datetime import datetime, timezone
+
+
+def parse_ts(value):
+    """Gitea emits RFC3339. Return an aware datetime, or None if it is unparseable.
+
+    Unparseable is NOT 'treat as fresh': the caller drops such a candidate, because a
+    timestamp we cannot read is a staleness check we cannot perform.
+    """
+    if not value:
+        return None
+    text = str(value).strip()
+    # Python < 3.11 cannot read a trailing 'Z' in fromisoformat.
+    if text.endswith(("Z", "z")):
+        text = text[:-1] + "+00:00"
+    try:
+        parsed = datetime.fromisoformat(text)
+    except ValueError:
+        return None
+    if parsed.tzinfo is None:
+        parsed = parsed.replace(tzinfo=timezone.utc)
+    return parsed
+
+
+def extract_reason(body, directive):
+    """Return the reason from the first line that STARTS with the directive, else None."""
+    pattern = re.compile(
+        r"^[ \t]*" + re.escape(directive) + r"(?=[ \t:]|$)[ \t:]*(.*)$"
+    )
+    for raw in (body or "").splitlines():
+        # A quoted line is somebody repeating an override, not casting one.
+        if raw.lstrip().startswith(">"):
+            continue
+        match = pattern.match(raw)
+        if match:
+            reason = match.group(1).strip()
+            return reason or None
+    return None
+
+
+def load_source(env_name, label):
+    """Return a list of items for one source, or [] with a stderr note if it is unusable."""
+    raw = os.environ.get(env_name)
+    if not raw or not raw.strip():
+        return []
+    try:
+        parsed = json.loads(raw)
+    except (ValueError, TypeError):
+        print(
+            "approval-override: %s is not valid JSON — ignoring that source" % label,
+            file=sys.stderr,
+        )
+        return []
+    if isinstance(parsed, list):
+        return parsed
+    # An error object, typically {"message": "token does not have ... scope"}. Say which,
+    # because "no override found" and "could not read the overrides" look identical from
+    # the outside and the first wasted a real debugging session.
+    detail = ""
+    if isinstance(parsed, dict):
+        detail = str(parsed.get("message") or parsed.get("errors") or "")[:200]
+    print(
+        "approval-override: %s is not a list (Gitea returned %s%s) — ignoring that source"
+        % (label, type(parsed).__name__, ": " + detail if detail else ""),
+        file=sys.stderr,
+    )
+    return []
+
+
+def candidates_from(items, source, directive, head_time):
+    """Yield (when, login, reason, raw_when, source) for every valid override in one source."""
+    found = []
+    for item in items:
+        if not isinstance(item, dict):
+            continue
+        login = ((item.get("user") or {}).get("login") or "").strip()
+        reason = extract_reason(item.get("body"), directive)
+        if not login or not reason:
+            continue
+        # A PENDING review is an unsubmitted private draft — not an act of anything.
+        state = str(item.get("state") or "").upper()
+        if source == "review" and state == "PENDING":
+            print(
+                "approval-override: dropping PENDING (unsubmitted) review by '%s'" % login,
+                file=sys.stderr,
+            )
+            continue
+        raw_when = (
+            item.get("submitted_at") or item.get("created_at") or item.get("updated_at") or ""
+        )
+        when = parse_ts(raw_when)
+        if when is None:
+            print(
+                "approval-override: dropping %s override by '%s' — unreadable timestamp %r"
+                % (source, login, raw_when),
+                file=sys.stderr,
+            )
+            continue
+        # An override must post-date the commit it covers.
+        if head_time is not None and when <= head_time:
+            print(
+                "approval-override: dropping STALE %s override by '%s' (%s) — not newer than "
+                "head commit (%s). Re-post it to override the current commit."
+                % (source, login, when.isoformat(), head_time.isoformat()),
+                file=sys.stderr,
+            )
+            continue
+        found.append((when, login, reason, raw_when, source))
+    return found
+
+
+def main():
+    directive = os.environ.get("DIRECTIVE", "/approve-override").strip()
+    head_time = parse_ts(os.environ.get("HEAD_TIME", ""))
+
+    found = []
+    found += candidates_from(
+        load_source("COMMENTS_JSON", "COMMENTS_JSON"), "comment", directive, head_time
+    )
+    found += candidates_from(
+        load_source("REVIEWS_JSON", "REVIEWS_JSON"), "review", directive, head_time
+    )
+
+    if not found:
+        return 0
+
+    # Latest wins, across both sources: a later override supersedes an earlier one.
+    when, login, reason, raw_when, source = max(found, key=lambda c: c[0])
+    # Keep the reason on one line — this becomes a TSV field.
+    reason = " ".join(reason.split())
+    print("%s\t%s\t%s\t%s" % (login, raw_when, source, reason))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -25,10 +25,45 @@
 # true: if this cannot evaluate, an unreviewed change to the gate itself could merge. A skip is a
 # BYPASS, not a gap. So: cannot determine -> BLOCK.
 #
+# THE ADMIN OVERRIDE (added 2026-10-01), and the lock-out it clears.
+#
+# The rule above was UNSATISFIABLE for the one human who runs this repo. Gitea refuses to let a
+# pull request's author approve their own PR, and `jakman85` is the only collaborator — so "an
+# APPROVED review from someone who is not its author" is a condition the sole administrator can
+# never meet. Every sensitive-path PR they opened was permanently unmergeable, INCLUDING the PR
+# that would fix it. Measured 2026-10-01: PR #35, gate green in 1m41s, zero reviews, blocked.
+#
+# So an administrator may waive the review by COMMENTING on the PR:
+#
+#     /approve-override <reason>
+#
+# Authority comes from the LIVE Gitea permission API, never from the comment — the commenter must
+# hold a permission listed in `.override.authorized_permissions` (admin/owner) at the moment the
+# gate runs, so an old comment from a since-demoted account is worth nothing. The directive must
+# start a line, a reason is mandatory, and the comment must POST-DATE the head commit: an override
+# earned at commit A does not cover commit B. Choosing the candidate comment is the job of the
+# pure, offline, unit-tested ci/approval-override.py; AUTHORISING it is the job of this stage,
+# which fails closed.
+#
+# WHAT THIS COSTS, STATED PLAINLY. For an administrator, "a change to the gate cannot be gated by
+# the gate" stops being structural and becomes advisory. That is the unavoidable consequence of
+# there being one human; the alternative is a repository its owner cannot change. What is bought
+# back is a RECORD — the waiver is an explicit act, attributed to an account, carrying a stated
+# reason, in the PR history and in this stage's output. A silent bypass keyed on "is the author an
+# admin?" would have been less code and is exactly the fail-open shape this repo keeps finding.
+#
+# AND THE PART THAT IS NOT COMFORTABLE. Authorisation is by permission ALONE, with no allowlist of
+# human accounts (operator decision, 2026-10-01, taken with this consequence on the table). The
+# AGENT account `gitea-admin` also reports permission `owner` here, so an agent can clear the
+# sensitive-path review on its own PR by posting the directive — and agents author most PRs in
+# this repo, so that is the common case, not an edge case. To narrow it without touching this
+# logic, add an `override_admins` list to infra/gitea/approval-policy.json: the check below
+# already requires membership whenever that key is present.
+#
 # EXIT CODES
-#   0  no sensitive paths touched (gate alone is sufficient), OR sensitive paths touched and a
-#      non-author APPROVED review is present
-#   1  sensitive paths touched without such a review, OR the policy could not be evaluated
+#   0  no sensitive paths touched (gate alone is sufficient), OR sensitive paths touched and
+#      either a non-author APPROVED review or a valid admin override is present
+#   1  sensitive paths touched with neither, OR the policy could not be evaluated
 set -euo pipefail
 CI_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 REPO_ROOT="$(cd "${CI_DIR}/.." && pwd)"
@@ -65,6 +100,17 @@ api() { curl -sS --max-time 20 -H "Authorization: token ${TOKEN}" "$@"; }
 FILES_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/pulls/${PR}/files?limit=1000" || true)"
 PR_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/pulls/${PR}" || true)"
 REVIEWS_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/pulls/${PR}/reviews" || true)"
+# PR conversation comments live under the ISSUES endpoint. `/pulls/{n}/comments` returns review
+# (diff) comments ONLY, so an override posted in the normal comment box would never appear in it.
+# Verified against this Gitea (1.24.3) rather than assumed.
+#
+# This one is NOT in the must-parse loop below, and that is deliberate. /issues/{n}/comments sits
+# behind Gitea's *issue* token scope, which `SETTINGS_API_TOKEN` does not carry (it is minted
+# `[read:repository, read:admin]`) — so it returns an error OBJECT, not an array. Blocking on that
+# turned "CI cannot read your override" into a message that said "you did not post one", and cost a
+# real debugging session on PR #38. It is handled explicitly, and loudly, further down: an
+# unreadable source can only cause a false BLOCK, never a false pass.
+COMMENTS_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/issues/${PR}/comments?limit=1000" || true)"
 for v in FILES_JSON PR_JSON REVIEWS_JSON; do
  jq -e . >/dev/null 2>&1 <<<"${!v}" || {
    err "could not read ${v} from Gitea for PR #${PR} — BLOCKING rather than assuming."
@@ -107,11 +153,122 @@ if [ -n "$APPROVER" ]; then
  exit 0
 fi

+# ─────────────────────────────────────────────────────────────────────────────────────────
+# No non-author approval exists. Before refusing, look for an administrator's explicit
+# override. See the header for why this hatch exists and what it costs.
+# ─────────────────────────────────────────────────────────────────────────────────────────
+OVERRIDE_ENABLED="$(jq -r '.override.enabled // false' "$DECL")"
+DIRECTIVE="$(jq -r '.override.comment_directive // "/approve-override"' "$DECL")"
+HEAD_SHA="$(jq -r '.head.sha // empty' <<<"$PR_JSON")"
+
+if [ "$OVERRIDE_ENABLED" = "true" ]; then
+  # The head commit's timestamp anchors staleness. Read it from GITEA, not from the local
+  # tree: CI checks out a merge commit, whose date is not the head commit's date.
+  HEAD_TIME=""
+  if [ -n "$HEAD_SHA" ]; then
+    HEAD_COMMIT_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/git/commits/${HEAD_SHA}" || true)"
+    HEAD_TIME="$(jq -r '.commit.committer.date // .created // empty' <<<"${HEAD_COMMIT_JSON:-null}" 2>/dev/null || true)"
+  fi
+  if [ -z "$HEAD_TIME" ]; then
+    err "could not read the head commit's timestamp for PR #${PR} — BLOCKING."
+    err "  Without it an override cannot be aged, and an unaged override is one that silently"
+    err "  covers commits nobody waived review on."
+    summary "- approval-policy: **BLOCKED — head commit timestamp unreadable** (cannot age an override)"
+    exit 1
+  fi
+
+  # Is the comments source actually readable? Say so plainly if not — a scope problem and "you
+  # never posted an override" are indistinguishable from the outside, and conflating them is what
+  # made PR #38 look like a code bug.
+  if ! jq -e 'type == "array"' >/dev/null 2>&1 <<<"${COMMENTS_JSON:-}"; then
+    SCOPE_MSG="$(jq -r '.message // empty' <<<"${COMMENTS_JSON:-null}" 2>/dev/null || true)"
+    warn "cannot read PR COMMENTS from Gitea${SCOPE_MSG:+ — ${SCOPE_MSG}}."
+    warn "  /issues/${PR}/comments needs the *issue* token scope; SETTINGS_API_TOKEN is minted"
+    warn "  [read:repository, read:admin] (infra/gitea/set-ci-token.sh). An override posted as an"
+    warn "  ordinary PR COMMENT is therefore invisible to CI — post it as a REVIEW of type"
+    warn "  'Comment' instead, which /pulls/${PR}/reviews exposes under read:repository."
+    summary "- approval-policy: note — PR comments unreadable (token lacks the issue scope); use a **review** to override"
+    COMMENTS_JSON='[]'
+  fi
+
+  # Pure, offline, unit-tested selection: which comment or review, if any, is a live override?
+  OVERRIDE_TSV="$(COMMENTS_JSON="$COMMENTS_JSON" REVIEWS_JSON="$REVIEWS_JSON" \
+    HEAD_TIME="$HEAD_TIME" DIRECTIVE="$DIRECTIVE" \
+    python3 "${CI_DIR}/approval-override.py")" || {
+      err "ci/approval-override.py failed — BLOCKING rather than assuming there is no override."
+      summary "- approval-policy: **BLOCKED — override parser failed**"
+      exit 1; }
+
+  if [ -n "$OVERRIDE_TSV" ]; then
+    OV_USER="$(cut -f1 <<<"$OVERRIDE_TSV")"
+    OV_WHEN="$(cut -f2 <<<"$OVERRIDE_TSV")"
+    OV_SRC="$(cut -f3 <<<"$OVERRIDE_TSV")"
+    OV_WHY="$(cut -f4- <<<"$OVERRIDE_TSV")"
+
+    # AUTHORITY IS MEASURED LIVE, never inferred from the comment or from a name in git.
+    PERM_JSON="$(api "${GITEA_URL}/api/v1/repos/${GITEA_REPO}/collaborators/${OV_USER}/permission" || true)"
+    OV_PERM="$(jq -r '.permission // empty' <<<"${PERM_JSON:-null}" 2>/dev/null || true)"
+    if [ -z "$OV_PERM" ]; then
+      err "'${OV_USER}' posted ${DIRECTIVE}, but their permission on ${GITEA_REPO} could not be"
+      err "  read — BLOCKING. An override whose authority is unverified is not an override."
+      summary "- approval-policy: **BLOCKED — cannot verify \`${OV_USER}\`'s permission**"
+      exit 1
+    fi
+
+    # Authorised when the LIVE permission is one of the declared ones and — only if an
+    # `override_admins` allowlist exists at all — the account is on it. No allowlist means
+    # permission alone, which is the operator's standing decision; adding the key narrows
+    # this with no change to any code.
+    AUTHORISED="$(DECL="$DECL" OV_PERM="$OV_PERM" OV_USER="$OV_USER" python3 -c '
+import json, os
+d = json.load(open(os.environ["DECL"])).get("override", {})
+perms = d.get("authorized_permissions") or ["admin", "owner"]
+allow = d.get("override_admins")
+ok_perm = os.environ["OV_PERM"] in perms
+ok_list = True if allow is None else os.environ["OV_USER"] in allow
+print("yes" if (ok_perm and ok_list) else "no")
+')"
+
+    if [ "$AUTHORISED" = "yes" ]; then
+      if [ "$OV_USER" = "$AUTHOR" ]; then
+        warn "ADMIN OVERRIDE — SELF-WAIVED. '${OV_USER}' authored this PR and waived the"
+        warn "  non-author review requirement on their own change. Permitted (permission"
+        warn "  '${OV_PERM}'), deliberate, and shouted about here precisely because it is not free."
+      else
+        warn "ADMIN OVERRIDE by '${OV_USER}' (permission '${OV_PERM}'); author is '${AUTHOR}'."
+      fi
+      warn "  via:    PR ${OV_SRC}"
+      warn "  when:   ${OV_WHEN}  (head commit ${HEAD_SHA:0:8}, ${HEAD_TIME})"
+      warn "  reason: ${OV_WHY}"
+      warn "  review waived on: $(printf '%s ' $SENSITIVE)"
+      ok "sensitive paths cleared by administrator override — this stage does not block."
+      SELFNOTE=""; [ "$OV_USER" = "$AUTHOR" ] && SELFNOTE=", **self-waived as author**"
+      summary "- approval-policy: **OVERRIDDEN by \`${OV_USER}\`** (permission \`${OV_PERM}\`, via PR ${OV_SRC}${SELFNOTE})"
+      summary "  - reason: ${OV_WHY}"
+      summary "  - review waived on: $(printf '\`%s\` ' $SENSITIVE)"
+      exit 0
+    fi
+
+    err "'${OV_USER}' posted ${DIRECTIVE} but is NOT authorised to override: live permission is"
+    err "  '${OV_PERM}'. The authorised set is declared in infra/gitea/approval-policy.json."
+    summary "- approval-policy: **BLOCKED — \`${OV_USER}\` may not override** (permission \`${OV_PERM}\`)"
+    exit 1
+  fi
+fi
+
 err "This PR changes the pipeline or the security posture it rests on, and has NO approving"
 err "  review from anyone other than its author ('${AUTHOR}')."
 err "  A change to the gate cannot be gated by the gate — that is the one thing CI cannot"
 err "  honestly approve for itself. Get a human review, or move the change out of:"
 printf '    %s\n' $SENSITIVE >&2
 err "  The list lives in infra/gitea/approval-policy.json, with a reason per entry."
-summary "- approval-policy: **BLOCKED — sensitive paths need a non-author approval**"
+if [ "$OVERRIDE_ENABLED" = "true" ]; then
+  err "  Or, if you hold admin/owner on this repository, post this on the PR:"
+  err "      ${DIRECTIVE} <reason>"
+  err "  as a REVIEW of type 'Comment' (not the plain comment box — CI's token cannot read"
+  err "  /issues/${PR}/comments without the issue scope; see infra/gitea/set-ci-token.sh)."
+  err "  The reason is mandatory, the directive must START a line, and it must be NEWER than"
+  err "  head commit ${HEAD_SHA:0:8} — push again and you must override again."
+fi
+summary "- approval-policy: **BLOCKED — needs a non-author approval, or an admin \`${DIRECTIVE}\` comment**"
 exit 1
@@ -1,5 +1,10 @@
 #!/usr/bin/env bash
-# ci/render-golden.sh — Stage: render==golden reproducibility gate.
+# ci/render-golden.sh — Stage: the renderer is trustworthy.
+#   (1) catalog/render.py passes its OWN output-contract self-test, and
+#   (2) every committed apps/ golden reproduces byte-for-byte from the canonical render.
+# Two independent assertions about one subject, catalog/render.py. (1) was added
+# 2026-10-01 because nothing in ci/ or .gitea/ had ever invoked `--selftest`; see the
+# block above the invocation for why it lives here and not in a conformance check.
 #
 # WHY: the committed apps/ tree is a set of GOLDENS produced by ONE canonical
 # invocation of catalog/render.py. If a golden is hand-edited, or a catalog
@@ -54,6 +59,54 @@ if [ ! -f "$RENDER_PY" ]; then
  summary "- render-golden: **SKIPPED** (renderer missing)"
  exit 3
 fi
+# --- The renderer's OWN self-test, before anything is compared ---------------
+# `catalog/render.py --selftest` is ~106 executed assertions on the renderer's output
+# contract (default-deny covers both directions, the Kyverno namespace hook label, a valid
+# auth-mode, the oauth2-proxy sidecar's cookie-name/-domain/-refresh flags, the
+# placeholder-substitution guard, the observability refusals...). Until now NOTHING in ci/
+# or .gitea/ invoked it — measured by grep on 2026-09-30 and recorded as the top follow-up
+# of decisions/2026-09-30-oauth2-proxy-cookie-collision-fleet.md — so every one of those
+# assertions was dev-time only and a merge gate could not tell a sound renderer from a
+# gutted one.
+#
+# WHY HERE, AND NOT IN A CONFORMANCE CHECK. The lead was
+# conformance/tests/l0-static/auditability-proofs-producible/check.sh, which drives
+# `conformance/lib/{record_run,status}.py --selftest` "because nothing else drives them".
+# That loop is NOT the right hook: that check serves R-SEC-03 and its subject is the
+# EVIDENCE CHAIN (can a forged or stale run become a proof?). The renderer is not part of
+# that chain, and hanging it there would make the check's name and its requirement mapping
+# a lie. This stage's subject already IS catalog/render.py — it exists to assert that the
+# renderer's output is trustworthy — so the renderer's own contract test belongs in it.
+#
+# It runs BEFORE the catalog/empty-catalog exit-3 below on purpose: the self-test is
+# deterministic and catalog-independent (it renders synthetic ServiceDefinitions in
+# memory), so an empty catalog must not be able to skip it.
+#
+# render-golden and the self-test are INDEPENDENT assertions, which is why both are needed:
+# render-golden catches golden-vs-renderer drift but is satisfied by ANY renderer whose
+# output was re-rendered, including one that just lost a contract invariant and had all ten
+# goldens regenerated in the same commit. The self-test catches exactly that.
+#
+# FAILURE IS A HARD FAIL, NOT A SKIP: the renderer is source in this repo, so "could not
+# run its self-test" has no legitimate environmental cause once python3 and render.py are
+# both present (both asserted above). ci/gate.sh maps this stage's rc 1 to BLOCKED.
+log "Stage: catalog/render.py --selftest (the renderer's own output-contract assertions)"
+selftest_out="$(mktemp)"
+if "$PYTHON" "$RENDER_PY" --selftest >"$selftest_out" 2>&1; then
+  ok "render.py --selftest passed"
+  summary "- render-selftest: **passed** (\`catalog/render.py --selftest\`)"
+else
+  err "catalog/render.py --selftest FAILED — the renderer violates its own output contract."
+  err "  Every golden under apps/ is produced by this renderer, so the goldens below cannot"
+  err "  be trusted even if they reproduce byte-for-byte. Reproduce with:"
+  err "    ${PYTHON} catalog/render.py --selftest"
+  sed 's/^/    /' "$selftest_out" >&2 || true
+  summary "- render-selftest: **FAILED** (\`catalog/render.py --selftest\` — renderer contract violated)"
+  rm -f "$selftest_out"
+  exit 1
+fi
+rm -f "$selftest_out"
+
 if [ ! -d "$CATALOG_DIR" ] || ! compgen -G "${CATALOG_DIR}/*.yaml" >/dev/null 2>&1; then
  skip "no catalog ServiceDefinitions under '${CATALOG_DIR}' — nothing to render."
  summary "- render-golden: **SKIPPED** (empty catalog)"
@@ -0,0 +1,232 @@
+#!/usr/bin/env bash
+# ci/tests/approval-override.test.sh — proves ci/approval-override.py selects the override it
+# claims to select, and REJECTS the ones it claims to reject.
+#
+# This parser decides whether a human review requirement gets waived. Every rule in it exists
+# because the obvious implementation ("does the body contain the directive?") is wrong in a way
+# that is invisible until abused: a quoted directive would re-arm an override, a bare directive
+# would waive review with no stated reason, and an override posted against commit A would go on
+# covering commit B forever. A guard is worth nothing until somebody has watched it refuse, which
+# is the same reasoning as ci/tests/decisions-log.test.sh one layer up.
+#
+# The parser is pure and offline BY DESIGN — authorisation (the live permission lookup) lives in
+# ci/approval-policy.sh, which cannot be tested without Gitea. That split is what makes this file
+# possible, and it is also why a green run here does NOT prove the stage authorises correctly.
+set -uo pipefail
+CI_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+SUT="${CI_DIR}/approval-override.py"
+
+PASS=0 FAIL=0
+ok()  { printf '  ok   %s\n' "$*"; PASS=$((PASS+1)); }
+bad() { printf '  FAIL %s\n' "$*"; FAIL=$((FAIL+1)); }
+
+HEAD_TIME="2026-10-01T12:00:00Z"
+AFTER="2026-10-01T12:30:00Z"
+BEFORE="2026-10-01T11:30:00Z"
+
+# run <comments-json> -> stdout of the parser (stderr is diagnostic only)
+run() {
+  COMMENTS_JSON="$1" REVIEWS_JSON="" HEAD_TIME="${2:-$HEAD_TIME}" DIRECTIVE="${3:-/approve-override}" \
+    python3 "$SUT" 2>/dev/null
+}
+
+# run_rev <reviews-json> — the same, but through the REVIEWS source.
+run_rev() {
+  COMMENTS_JSON="" REVIEWS_JSON="$1" HEAD_TIME="${2:-$HEAD_TIME}" DIRECTIVE="${3:-/approve-override}" \
+    python3 "$SUT" 2>/dev/null
+}
+
+# comment <login> <created_at> <body>
+comment() {
+  python3 -c '
+import json, sys
+print(json.dumps([{"user": {"login": sys.argv[1]}, "created_at": sys.argv[2], "body": sys.argv[3]}]))
+' "$1" "$2" "$3"
+}
+
+expect_none() {
+  local label="$1" out
+  out="$(run "$2")"
+  if [ -z "$out" ]; then ok "$label"; else bad "$label (got: ${out})"; fi
+}
+
+expect_user() {
+  local label="$1" json="$2" want="$3" out got
+  out="$(run "$json")"
+  got="$(cut -f1 <<<"$out")"
+  if [ "$got" = "$want" ]; then ok "$label"; else bad "$label (wanted ${want}, got '${got}')"; fi
+}
+
+expect_reason() {
+  local label="$1" json="$2" want="$3" out got
+  out="$(run "$json")"
+  got="$(cut -f4- <<<"$out")"
+  if [ "$got" = "$want" ]; then ok "$label"; else bad "$label (wanted '${want}', got '${got}')"; fi
+}
+
+# check <label> <actual> <wanted>
+check() {
+  if [ "$2" = "$3" ]; then ok "$1"; else bad "$1 (wanted '${3}', got '${2}')"; fi
+}
+
+# review <login> <submitted_at> <state> <body>
+review() {
+  python3 -c '
+import json, sys
+print(json.dumps([{"user": {"login": sys.argv[1]}, "submitted_at": sys.argv[2],
+                   "state": sys.argv[3], "body": sys.argv[4]}]))
+' "$1" "$2" "$3" "$4"
+}
+
+echo "== ci/approval-override.py"
+
+# --- the happy path -------------------------------------------------------------------
+expect_user "a directive with a reason, posted after the head commit, is selected" \
+  "$(comment jakman85 "$AFTER" '/approve-override sole admin, gate is green')" jakman85
+
+expect_reason "the reason is returned verbatim" \
+  "$(comment jakman85 "$AFTER" '/approve-override sole admin, gate is green')" \
+  'sole admin, gate is green'
+
+# --- absence is a normal answer, not an error -----------------------------------------
+expect_none "no comments at all" '[]'
+expect_none "an ordinary comment" "$(comment jakman85 "$AFTER" 'looks good to me')"
+
+# --- a reason is MANDATORY ------------------------------------------------------------
+expect_none "a bare directive (no reason) is refused" \
+  "$(comment jakman85 "$AFTER" '/approve-override')"
+expect_none "a directive followed by only whitespace is refused" \
+  "$(comment jakman85 "$AFTER" '/approve-override    ')"
+
+# --- the directive must START a line --------------------------------------------------
+expect_none "the directive mid-sentence does not count" \
+  "$(comment jakman85 "$AFTER" 'you could just /approve-override this one')"
+expect_none "a QUOTED directive does not re-arm an override" \
+  "$(comment jakman85 "$AFTER" '> /approve-override quoting someone else')"
+expect_user "leading whitespace before the directive is fine" \
+  "$(comment jakman85 "$AFTER" '   /approve-override indented but mine')" jakman85
+expect_user "the directive on a later line of a longer comment counts" \
+  "$(comment jakman85 "$AFTER" 'some preamble
+/approve-override the real reason here')" jakman85
+
+# --- staleness: an override must post-date the commit it waives review on -------------
+expect_none "an override OLDER than the head commit is dropped as stale" \
+  "$(comment jakman85 "$BEFORE" '/approve-override stale, predates the push')"
+expect_none "an override exactly AT the head commit time is dropped (strictly newer)" \
+  "$(comment jakman85 "$HEAD_TIME" '/approve-override same instant')"
+
+# --- malformed input is never a free pass --------------------------------------------
+expect_none "an unreadable timestamp drops the candidate" \
+  "$(comment jakman85 'not-a-date' '/approve-override bad clock')"
+expect_none "a comment with no author is dropped" \
+  '[{"user": {}, "created_at": "2026-10-01T12:30:00Z", "body": "/approve-override no login"}]'
+expect_none "invalid JSON yields no override (and never a traceback into stdout)" \
+  'this is not json'
+expect_none "a JSON object instead of a list yields no override" \
+  '{"user": "nope"}'
+
+# --- latest wins ----------------------------------------------------------------------
+TWO='[
+  {"user": {"login": "early"},  "created_at": "2026-10-01T12:10:00Z", "body": "/approve-override first"},
+  {"user": {"login": "latest"}, "created_at": "2026-10-01T12:40:00Z", "body": "/approve-override second"}
+]'
+expect_user "the LATEST override supersedes an earlier one" "$TWO" latest
+expect_reason "and brings its own reason" "$TWO" second
+
+# --- offsets, not just Z --------------------------------------------------------------
+# Gitea emits RFC3339 WITH AN OFFSET (measured: 2026-10-01T23:09:08-05:00), not a trailing Z.
+# A parser that only handles 'Z' silently drops every real override, so this case is load-bearing.
+expect_user "an RFC3339 offset timestamp (-05:00), as Gitea actually emits, parses" \
+  '[{"user": {"login": "jakman85"}, "created_at": "2026-10-01T08:30:00-05:00", "body": "/approve-override offset form"}]' \
+  jakman85
+expect_none "and an offset timestamp that is genuinely earlier is still stale" \
+  '[{"user": {"login": "jakman85"}, "created_at": "2026-10-01T06:00:00-05:00", "body": "/approve-override offset, stale"}]'
+
+# --- the directive is configurable ----------------------------------------------------
+out="$(run "$(comment jakman85 "$AFTER" '/admin-waive because policy says so')" "$HEAD_TIME" '/admin-waive')"
+if [ "$(cut -f1 <<<"$out")" = "jakman85" ]; then
+  ok "a custom directive from the declaration is honoured"
+else
+  bad "a custom directive from the declaration is honoured (got '${out}')"
+fi
+expect_none "the default directive does not match a custom-directive comment" \
+  "$(comment jakman85 "$AFTER" '/admin-waive because policy says so')"
+
+# --- no staleness anchor: do not invent one -------------------------------------------
+out="$(COMMENTS_JSON="$(comment jakman85 "$BEFORE" '/approve-override no anchor')" \
+       HEAD_TIME="" DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+if [ "$(cut -f1 <<<"$out")" = "jakman85" ]; then
+  ok "with no HEAD_TIME the staleness test is skipped, not faked (caller blocks instead)"
+else
+  bad "with no HEAD_TIME the candidate should still be returned (got '${out}')"
+fi
+
+# --- the REVIEWS source ---------------------------------------------------------------
+# Load-bearing, not redundant. Measured on PR #38: SETTINGS_API_TOKEN is minted
+# [read:repository, read:admin], /issues/{n}/comments needs the ISSUE scope, so a well-formed
+# override in the comment box came back as an error object and was invisible. /pulls/{n}/reviews
+# IS readable with that token, so a review body is the route that works today.
+check "a review body carries an override" \
+  "$(cut -f1 <<<"$(run_rev "$(review jakman85 "$AFTER" COMMENT '/approve-override via a review')")")" \
+  jakman85
+check "the source is reported as 'review'" \
+  "$(cut -f3 <<<"$(run_rev "$(review jakman85 "$AFTER" COMMENT '/approve-override via a review')")")" \
+  review
+check "the source is reported as 'comment' for a comment" \
+  "$(cut -f3 <<<"$(run "$(comment jakman85 "$AFTER" '/approve-override via a comment')")")" \
+  comment
+
+out="$(run_rev "$(review jakman85 "$BEFORE" COMMENT '/approve-override stale review')")"
+check "a stale review override is dropped too" "${out:-<none>}" "<none>"
+
+out="$(run_rev "$(review jakman85 "$AFTER" PENDING '/approve-override unsubmitted draft')")"
+check "a PENDING (unsubmitted) review is NOT an act of anything" "${out:-<none>}" "<none>"
+
+check "an APPROVED review carrying the directive still counts as an override" \
+  "$(cut -f1 <<<"$(run_rev "$(review jakman85 "$AFTER" APPROVED '/approve-override belt and braces')")")" \
+  jakman85
+
+out="$(run_rev "$(review jakman85 "$AFTER" COMMENT 'ordinary review prose, no directive')")"
+check "an ordinary review body is not an override" "${out:-<none>}" "<none>"
+
+# --- an unreadable source degrades to 'no candidates', never to a pass ----------------
+# The real-world shape: Gitea answers with {"message": "token does not have ... scope"}.
+ERR_OBJ='{"message":"token does not have at least one of required scope(s): [read:issue]"}'
+out="$(COMMENTS_JSON="$ERR_OBJ" REVIEWS_JSON="" HEAD_TIME="$HEAD_TIME" \
+       DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+check "an error OBJECT from the comments API yields no override" "${out:-<none>}" "<none>"
+
+out="$(COMMENTS_JSON="$ERR_OBJ" \
+       REVIEWS_JSON="$(review jakman85 "$AFTER" COMMENT '/approve-override reviews still work')" \
+       HEAD_TIME="$HEAD_TIME" DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+check "an unreadable COMMENTS source does not suppress a valid REVIEW override" \
+  "$(cut -f1 <<<"$out")" jakman85
+
+# and it must SAY so, or a scope problem reads as 'you never posted one' (it did, on PR #38)
+errout="$(COMMENTS_JSON="$ERR_OBJ" REVIEWS_JSON="" HEAD_TIME="$HEAD_TIME" \
+          DIRECTIVE=/approve-override python3 "$SUT" 2>&1 >/dev/null)"
+case "$errout" in
+  *"not a list"*read:issue*) ok "the diagnostic names the scope Gitea complained about" ;;
+  *) bad "the diagnostic should quote Gitea's message (got: ${errout})" ;;
+esac
+
+# --- latest wins ACROSS the two sources ----------------------------------------------
+out="$(COMMENTS_JSON="$(comment early "2026-10-01T12:10:00Z" '/approve-override from a comment')" \
+       REVIEWS_JSON="$(review later "2026-10-01T12:50:00Z" COMMENT '/approve-override from a review')" \
+       HEAD_TIME="$HEAD_TIME" DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+check "the latest override wins across both sources (review newer)" "$(cut -f1 <<<"$out")" later
+check "and its source is labelled" "$(cut -f3 <<<"$out")" review
+
+out="$(COMMENTS_JSON="$(comment later "2026-10-01T12:50:00Z" '/approve-override from a comment')" \
+       REVIEWS_JSON="$(review early "2026-10-01T12:10:00Z" COMMENT '/approve-override from a review')" \
+       HEAD_TIME="$HEAD_TIME" DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+check "the latest override wins across both sources (comment newer)" "$(cut -f1 <<<"$out")" later
+check "and that source is labelled too" "$(cut -f3 <<<"$out")" comment
+
+# --- both sources absent entirely -----------------------------------------------------
+out="$(COMMENTS_JSON="" REVIEWS_JSON="" HEAD_TIME="$HEAD_TIME" \
+       DIRECTIVE=/approve-override python3 "$SUT" 2>/dev/null)"
+check "no sources at all yields no override, and no crash" "${out:-<none>}" "<none>"
+
+printf '\n  %d passed, %d failed\n' "$PASS" "$FAIL"
+[ "$FAIL" -eq 0 ] || exit 1
@@ -42,7 +42,16 @@ cp "${REAL_CI_DIR}/gate.sh" "${REAL_CI_DIR}/lib.sh" "${REAL_CI_DIR}/versions.env
 # 15 of 47 assertions failed on `main` for that reason alone, including every one that expects a
 # PASS/UNDETERMINED/NOT-APPLICABLE headline. The suite was measuring the missing stub, not the
 # decision logic. `decisions-log` is added here in the same change that adds it to gate.sh.
-for stage in spec-immutability decisions-log lint l0 render-golden l1 gitea-settings conformance; do
+#
+# AND IT HAPPENED AGAIN, with `approval-policy` (fixed 2026-10-01). Same shape, same cost: the
+# stage was added to gate.sh by PR #17, never added here, so 15 of 50 assertions failed on `main`
+# — measured on 0c7e8dd before touching anything, to be sure it was not the change under test.
+# The note above was written about the previous instance and did not prevent the next one, which
+# says the comment is not the fix. THE FIX IS TO DERIVE THE LIST: it now comes from gate.sh's own
+# `run_stage` calls, so a stage added there can no longer be silently missing here.
+mapfile -t STAGES < <(sed -n 's/^run_stage[[:space:]]\+\([a-z0-9-]\+\).*/\1/p' "${REAL_CI_DIR}/gate.sh")
+[ "${#STAGES[@]}" -ge 8 ] || { echo "could not parse stage names out of gate.sh (got ${#STAGES[@]})" >&2; exit 1; }
+for stage in "${STAGES[@]}"; do
  cat > "${CI}/${stage}.sh" <<'STUB'
 #!/usr/bin/env bash
 n="$(basename "$0" .sh)"; v="STAGE_RC_${n//-/_}"
@@ -100,7 +100,8 @@ data:
  # platform/keycloak/postgres-cluster.yaml but is rendered by platform-config, which is the ONLY
  # tier that lists it), crafty (CRAFTY_L4_VIP), apps-wave-0 (nextcloud-db + semaphore-db:
  # S3_BACKUP_ENDPOINT, CNPG_BACKUP_PATH; pihole + pihole-dns: PIHOLE_DNS_VIP, LAN_CIDR, plus
-  # GATEWAY_VIP/UNIFI_L4_VIP/CRAFTY_L4_VIP read BY pihole to author its LAN split-horizon answers),
+  # GATEWAY_VIP/UNIFI_L4_VIP/CRAFTY_L4_VIP read BY pihole to author its LAN split-horizon answers;
+  # plex-remote: PLEX_L4_VIP; plex: PLEX_ADVERTISE_URL),
  # velero (S3_BACKUP_ENDPOINT, S3_BACKUP_REGION,
  # VELERO_BACKUP_BUCKET, VELERO_BACKUP_PREFIX), apps-vms (UNIFI_L4_VIP). A ${TOKEN} in an
  # UNSUBSTITUTED path would ship the literal string into the cluster.
@@ -175,6 +176,45 @@ data:
  # dev  -> 192.168.1.56  (base default; next free after crafty .55 in dev's .53-.60 pool)
  # prod -> 192.168.2.23  (next free after crafty .22 in prod's .20-.30 pool; see clusters/prod)
  PIHOLE_DNS_VIP: "192.168.1.56"             # PiHole: authoritative LAN DNS for ${BASE_DOMAIN}
+  # Plex's remote-access listener: raw tcp/32400 for REMOTE Plex clients (apps/wave-0/plex-remote.yaml).
+  # A FOURTH pinned L4 VIP, for the usual reason — the `shared` Gateway is an L7 HTTPS terminator and
+  # cannot carry Plex's own TLS. Plex clients do not resolve `plex.${BASE_DOMAIN}`; they fetch this
+  # server's address from plex.tv, so the Gateway's L7 path (which still serves the web UI, and was
+  # measured working) cannot put the server on a client's connection list at all.
+  #
+  # PINNED, AND THE PIN IS LOAD-BEARING OFF-REPO: the SOHO router forwards WAN tcp/32400 to this
+  # address by hand. A VIP taken from the pool head would silently move on a reschedule and break
+  # remote access with nothing in-cluster looking wrong. See runbooks/plex-remote-access.md.
+  #
+  # dev  -> 192.168.1.57  (base default; next free after pihole-dns .56 in dev's .53-.60 pool)
+  # prod -> 192.168.2.24  (next free after pihole-dns .23 in prod's .20-.30 pool; see clusters/prod)
+  PLEX_L4_VIP: "192.168.1.57"                # Plex: tcp/32400 remote access (NOT the web UI, which is L7)
+
+  # PLEX_ADVERTISE_URL — the connection URL Plex publishes to plex.tv ("Custom server access URLs",
+  # the ADVERTISE_IP env in catalog/plex.yaml, which the image writes into the `customConnections`
+  # preference on every pod start). Read catalog/plex.yaml's ADVERTISE_IP block for the full why;
+  # the short version is the only thing that matters when editing this line:
+  #
+  #   IT MUST BE A `<dashed-ip>.<serverhash>.plex.direct` NAME, NEVER A BARE IP. Plex serves 32400
+  #   with a `*.<serverhash>.plex.direct` wildcard certificate and NO IP SAN, so `https://<ip>:32400`
+  #   fails TLS validation for every client. That exact mistake is what broke this service on
+  #   2026-10-04: with no completable connection, plex.tv published nothing, the server vanished from
+  #   the account and clients reported "not authorized". The dashed name is public DNS that resolves
+  #   to the embedded address, so it points at the LAN VIP AND matches the cert.
+  #
+  # WHY IT IS A WHOLE URL AND NOT BUILT FROM ${PLEX_L4_VIP}: postBuild substitution is plain
+  # envsubst — it cannot turn 192.168.2.24 into 192-168-2-24 — and `<serverhash>` is this server's
+  # certificate identity, derivable from nothing in this tree. Re-read it from the live cert
+  # (`openssl s_client -connect <vip>:32400 | openssl x509 -noout -subject`) if Plex ever reissues.
+  #
+  # The dashed IP here MUST track ${PLEX_L4_VIP} above. They are two halves of one address and
+  # nothing checks that they agree.
+  #
+  # dev  -> PLACEHOLDER, non-functional. dev is decommissioned and has no claimed Plex server, so
+  #         there is no real hash to put here; the base default only has to be a defined, correctly
+  #         shaped string (StrictPostBuildSubstitutions makes an UNDEFINED var a hard failure).
+  # prod -> the real 192.168.2.24 + live server hash; see clusters/prod/kustomization.yaml.
+  PLEX_ADVERTISE_URL: "https://192-168-1-57.0000000000000000000000000000000f.plex.direct:32400"

  # The FLAT LAN this cluster's nodes and LB VIPs live on, in CIDR form. Consumed as the ingress
  # peer of the PiHole DNS CiliumNetworkPolicy (apps/wave-0/pihole-dns.yaml): a resolver reachable
@@ -0,0 +1,104 @@
+---
+# Flux Kustomization: Intel GPU device plugin — reconciles ./platform/intel-gpu-plugin.
+# Makes the three prod nodes advertise `gpu.intel.com/i915`, which is the resource a later PR will
+# have `hardware.gpu: intel-quicksync` request in place of today's non-working /dev/dri hostPath.
+#
+# WHY ITS OWN KUSTOMIZATION, AND NOT ONE LINE IN platform/kustomization.yaml
+#
+# The cheap option was a line in the shared `platform` tier plus a new platform/ directory, and
+# neither path is sensitive in infra/gitea/approval-policy.json, so it would have needed no human
+# reviewer on the policy's terms. It is still the wrong place, for the reason the repo already
+# wrote down in platform/kustomization.yaml when Cilium was split out of that tier:
+#
+#   * `platform` is `wait: true` over eleven-plus components, so its Ready condition cannot tell
+#     you "the GPU plugin is wedged" from "longhorn is slow".
+#   * The only emergency valve would be `flux suspend kustomization platform` — freezing storage,
+#     TLS, DNS, ingress, monitoring, Keycloak and KubeVirt in order to stop touching a device
+#     plugin.
+#   * The tier's 10m interval / 5m timeout are explicitly called out there as WRONG for a 3-node
+#     DaemonSet rollout. This is a 3-node DaemonSet rollout.
+#
+# And the consequence is worse here than for an average component, because of what sits downstream:
+# `platform` gates platform-config, which gates apps-wave-0 (wait:true), which gates apps-wave-1.
+# A plugin DaemonSet that never went healthy inside that tier would stall every app on "dependency
+# not ready" — i.e. a component added to fix a SILENT NO-OP (software transcoding) would have
+# acquired the ability to cause a REAL OUTAGE. That trade is the whole argument.
+#
+# The usual objection to this placement is the approval cost, and it does not apply: Gitea enforces
+# `required_approvals: 1` with `block_admin_merge_override: true` on EVERY PR here regardless of
+# path, and authors cannot self-approve. So this change needed the operator's approval either way;
+# touching clusters/** costs nothing extra and buys a blast radius of exactly one component.
+#
+# dependsOn: [cilium], WHICH IS THE MINIMUM AND ALSO THE EARLIEST
+#
+# Not platform-config (what headlamp/crafty/homepage/oauth2-proxy use) — those are leaf UIs that
+# genuinely need Keycloak + wildcard TLS + split-horizon DNS + Gateway. This plugin needs none of
+# it. It never uses the network and never talks to the API server; it talks to the local kubelet
+# over a unix socket. Its only real prerequisite is a working CNI, so its pod can get a sandbox at
+# all, and `cilium` already depends on `flux-system` (the root Kustomization that applies the
+# cluster-vars ConfigMap this one substitutes from), so that edge is transitively satisfied too.
+#
+# Depending on `cilium` rather than `platform` also gets the resource advertised EARLIEST — this
+# reconciles in parallel with the `platform` tier instead of behind all eleven of its components.
+# That matters on a cold start once the consumer PR lands: a Plex pod requesting a resource no node
+# advertises is Unschedulable, and apps-wave-0 is wait:true. Reconciling early shrinks that window
+# instead of widening it. It is not a GUARANTEE — see the ordering note below, which is a real
+# constraint on the next PR and not a formality.
+#
+# Kyverno (in `platform`) is deliberately NOT a dependency: every policy that could match this
+# workload is scoped to namespaces labelled app-namespace=true, and this namespace is labelled
+# "false" (platform/intel-gpu-plugin/namespace.yaml). So there is no admission race to lose.
+#
+# wait: true — A HEALTH SIGNAL, NOT A GATE. This is the one place this component departs from the
+# headlamp/crafty leaf template, and the departure is deliberate. Those run wait:false because a
+# leaf UI can be legitimately not-Ready for a long time during a cold start while it waits on
+# Keycloak/DNS/TLS, so its health is not a useful signal. This DaemonSet has no such dependency: it
+# reads sysfs and opens a socket, so on healthy nodes it should be Ready in seconds, and
+# not-Ready genuinely means BROKEN. wait:true turns `flux get kustomization intel-gpu-plugin` into
+# a one-line answer to "did the plugin come up on all three nodes", which is exactly the question
+# the operator must answer before the consumer PR may merge.
+#
+# NOTHING IN THE GRAPH dependsOn THIS, so wait:true gates nothing — the blast radius of an unhealthy
+# plugin is this Kustomization's own Ready condition and nothing else. Anyone later adding
+# `dependsOn: [intel-gpu-plugin]` is choosing to make plugin health gate their thing; for the app
+# waves that would hand a device plugin the power to stall every app, which is the thing this file
+# exists to prevent. Think before adding that edge.
+#
+# ORDERING — DO NOT REORDER THE STACK. The renderer change (catalog/render.py emitting
+# `gpu.intel.com/i915: 1` for `hardware.gpu: intel-quicksync` instead of the hostPath) must NOT
+# land until this plugin is merged AND confirmed advertising the resource on all three nodes. If
+# the renderer requests a resource no node advertises, the Plex pod is Unschedulable; apps-wave-0 is
+# wait:true and apps-wave-1 dependsOn it, so that stalls the whole app tier — turning today's silent
+# no-op into a real outage. This PR's only job is to make the resource exist.
+apiVersion: kustomize.toolkit.fluxcd.io/v1
+kind: Kustomization
+metadata:
+  name: intel-gpu-plugin
+  namespace: flux-system
+spec:
+  # Static vendored manifests with nothing to re-render; kustomize-controller is source-triggered,
+  # so this is only the drift-correction floor.
+  interval: 10m
+  retryInterval: 2m
+  # 10m, not the 5m the `platform` tier uses: a 3-node DaemonSet rollout under `wait: true` can
+  # exceed 5m on a cold start (image pull on three nodes), and the Cilium split-out comment calls
+  # out that exact mismatch as one of its reasons for leaving the shared tier.
+  timeout: 10m
+  prune: true
+  # See the header: a health SIGNAL for a component nothing depends on, not a gate.
+  wait: true
+  dependsOn:
+    - name: cilium   # the CNI — the only real prerequisite; transitively gives us flux-system too
+  sourceRef:
+    kind: GitRepository
+    name: flux-system
+  path: ./platform/intel-gpu-plugin
+  # Retained for parity with every other tier / future overlay knobs. This component references no
+  # ${VARS} today, and its manifests contain no `$` literals for envsubst to eat (unlike
+  # clusters/base/runbooks.yaml, which omits substitution for exactly that reason).
+  postBuild:
+    substituteFrom:
+      - kind: ConfigMap
+        name: cluster-vars
+  # No SOPS decryption patch: this component carries no Secret, so clusters/dev adds nothing for it
+  # (same as homepage/runbooks, unlike headlamp/crafty/oauth2-proxy/backup).
@@ -12,6 +12,15 @@ resources:
  # interval/timeout, and `flux suspend kustomization cilium` as an emergency valve that does not
  # freeze the other eleven platform components. See clusters/base/cilium.yaml.
  - cilium.yaml          # Flux Kustomization -> ./platform/cilium (dependsOn flux-system, wait:true)
+  # Intel GPU device plugin — its OWN Kustomization (dependsOn cilium, wait:true), NOT a line in
+  # platform/kustomization.yaml. Same reasoning as the Cilium split-out, applied to a leaf: inside
+  # that wait:true tier a wedged 3-node DaemonSet would be indistinguishable from eleven unrelated
+  # components AND would stall platform-config + every apps wave on "dependency not ready" — so a
+  # component added to fix a silent no-op (Plex software transcoding) could cause a real outage.
+  # wait:true here is a HEALTH SIGNAL, not a gate: nothing in the graph dependsOn it.
+  # It advertises `gpu.intel.com/i915`; the renderer change that REQUESTS that resource must land
+  # strictly later. See clusters/base/intel-gpu-plugin.yaml.
+  - intel-gpu-plugin.yaml # Flux Kustomization -> ./platform/intel-gpu-plugin (dependsOn cilium)
  - platform.yaml        # Flux Kustomization -> ./platform (CONTROLLERS: HelmReleases / CRDs)
  - platform-config.yaml # Flux Kustomization -> ./platform-config (CONFIGS: CRs; dependsOn platform)
  # SECOND-ORDER CDI configs — CRs whose CRD is created by a controller AT RUNTIME, not by any
@@ -476,6 +476,36 @@ patches:
      - op: replace
        path: /data/PIHOLE_DNS_VIP
        value: "192.168.2.23"
+      # Plex's remote-access listener VIP — raw tcp/32400 for REMOTE Plex clients
+      # (apps/wave-0/plex-remote.yaml). 192.168.2.24 is the NEXT FREE address in the prod pool
+      # 192.168.2.20-.30, verified live 2026-10-03: the CiliumLoadBalancerIPPool `default-pool`
+      # reported IPsTotal 11 / IPsUsed 4 / IPsAvailable 7, and the only four allocations are .20
+      # (cilium-gateway-shared), .21 (unifi-l4-tcp/-udp), .22 (crafty-l4-tcp/-udp) and .23
+      # (pihole-dns-tcp/-udp). It is inside the pool block the operator already asserted sits outside
+      # the SOHO router's DHCP scope, and it is not a node address (.10/.11/.12) or the
+      # control-plane VIP (.9).
+      #
+      # UNLIKE .23, THIS ONE IS ALSO A WAN PORT-FORWARD TARGET. The router forwards WAN tcp/32400
+      # here, by hand, off-repo — so changing this value is a two-place edit and the router is the
+      # half that CI cannot check. runbooks/plex-remote-access.md is the procedure.
+      #
+      # NOT ARP-PROBED. .23 was verified free with `arping` from a LAN host before it was claimed;
+      # for .24 only the in-cluster pool accounting above was checked, so an off-cluster squatter on
+      # this address would not have been caught. Probe before trusting it if remote access comes up
+      # with a conflict.
+      - op: replace
+        path: /data/PLEX_L4_VIP
+        value: "192.168.2.24"
+      # PLEX_ADVERTISE_URL — the prod VIP above in Plex's `plex.direct` form, which is the ONLY form
+      # that validates: the server's cert is `*.130072f20c334d6c8f63d0dfef87309b.plex.direct` with no
+      # IP SAN (read off the live listener 2026-10-04), so a bare `https://192.168.2.24:32400` fails
+      # TLS for every client — the 2026-10-04 outage. Measured with full verification, no `curl -k`:
+      # this URL answers /identity with 200 and TLS verify OK; the bare-IP form fails to connect.
+      # THE DASHED IP MUST MATCH /data/PLEX_L4_VIP ABOVE (192.168.2.24 -> 192-168-2-24); nothing
+      # enforces that. The hash is server identity, not a secret, and changes only if Plex reissues.
+      - op: replace
+        path: /data/PLEX_ADVERTISE_URL
+        value: "https://192-168-2-24.130072f20c334d6c8f63d0dfef87309b.plex.direct:32400"
      # The prod flat LAN. This is the 192.168.2.0/24 segment the SOHO ROUTER at 192.168.2.1 owns —
      # NOT pfSense's 192.168.1.0/24, which is a different network and not in the LAN-DNS path.
      # It scopes the ingress allow for PiHole's DNS listener to LAN clients only.
@@ -0,0 +1,134 @@
+# =============================================================================
+# L0 STATIC — `--api-route` is PER APP, ANCHORED, and DECLARED
+# Catalog: docs/14 L0 row ("Pre-apply, in CI") — runs WITHOUT a cluster.
+#
+# THE DEFECT THIS CLOSES. oauth2-proxy answers an unauthenticated request with a 302 to
+# Keycloak. Its built-in AJAX guard would send a 401 instead, but it fires only on an EXACT
+# `Accept: application/json`, which these apps' JavaScript does not consistently send. So a
+# browser XHR gets a 302, FOLLOWS IT INVISIBLY, has it satisfied silently by the live Keycloak
+# SSO session, and resolves 200 carrying the app's HTML where the JavaScript expected JSON — no
+# prompt, no console error, nothing logged as a failure. On a WRITE it is worse: a browser
+# following a 302 re-issues a POST as a GET and DROPS THE BODY, so the operator's change is
+# discarded and the UI reports success.
+#
+# MEASURED, twice, on two different apps and two different paths:
+#   jackett 2026-09-30 — on /api/ paths: 33 x 302 against 2 x 401, and 4 of 6 POSTs answered BY
+#           THE PROXY, never reaching Jackett (access-log upstream field `-`, not `/`). That is
+#           the operator's whole report: "GET works, most PUT/POST to add indexers fail,
+#           reloading never fixes it, and I am never re-prompted for login".
+#   radarr  2026-10-01, 23h after the cookie fix rolled (pod radarr-7d55cbffb9-42hmh, 0
+#           restarts) — 232 unauthenticated requests in the pod's whole lifetime and ALL 232 are
+#           the same call: `POST /signalr/messages/negotiate` from a browser UA, 302, upstream
+#           `-`. A left-open tab whose session the cookie rename orphaned, polling SignalR
+#           forever, silently redirected every time, NEVER re-prompting. The reported symptom,
+#           still running in production, on a service nobody was watching.
+#
+# WHY IT CANNOT BE A RENDERER CONSTANT — AND WHY THAT NEEDS A GATE. The path shape belongs to
+# the APP, not to the forward-auth pattern: radarr's broken path is `/signalr/`, jackett's is
+# `/api/`, and PIHOLE SERVES NO `/api/` AT ALL. One function renders all five sidecars
+# (catalog/render.py::_oauth2_sidecar_container), so the cheap fix is a constant — and a
+# constant applies one app's measurement to four apps nobody measured. The gate that already
+# exists cannot see that mistake: `ci/render-golden.sh` compares the render to the goldens, and
+# a constant plus regenerated goldens AGREE WITH EACH OTHER while both being wrong. Hence
+# assertion (2) below, which compares the rendered flag against each service's OWN
+# ServiceDefinition rather than against the renderer's output.
+#
+# WHY BREADTH IS THE DANGEROUS DIRECTION, AND IT IS NOT SYMMETRIC. A MISSING --api-route is a
+# silent 302 — bad, invisible, recoverable. A TOO BROAD one answers a plain PAGE LOAD with 401:
+# a logged-out operator gets a bare error page and is NEVER OFFERED A SIGN-IN, so the service
+# looks permanently broken and no reload recovers it. That is a worse outcome than the defect
+# the flag fixes, which is why the anchoring rule is asserted on three independent surfaces
+# (schema items.pattern, the renderer, and this check) and why assertion (5) fails if they drift
+# apart — if they do, the loosest one wins silently.
+#
+# GIVEN every oauth2-proxy sidecar this repo DEPLOYS (a container in any manifest at any depth:
+#       the rendered sidecars in apps/, crafty's hand-authored one in platform/), the
+#       ServiceDefinitions in catalog/ that generate five of them, and the renderer itself,
+# WHEN  each sidecar's `--api-route` values are collected (the flag is REPEATABLE, so all of
+#       them, in order) and each onboarded forward-auth service is paired with its golden,
+# THEN  (1) every rendered --api-route is anchored at `^/` and continues with literal path
+#           characters — `^/`, `^/.*`, `.*`, an unanchored `/api/` and any alternation FAIL;
+#       (2) every onboarded forward-auth service's rendered route list EQUALS its own
+#           ServiceDefinition's `auth.apiRoutes`, in order — which fails both a hand-edited
+#           golden and a renderer constant, neither of which render-golden can see;
+#       (3) BOTH cases are live in this tree: at least one service declares routes and at least
+#           one declares none and renders none (the pihole case), so neither path is an
+#           unexercised seam;
+#       (4) `_oauth2_sidecar_container()` reads `auth.apiRoutes` and contains NO literal path
+#           regex — a constant is the defect, back in the one place that reintroduces it
+#           fleet-wide in a single edit;
+#       (5) the anchoring pattern in the renderer and in the schema are the SAME pattern this
+#           check enforces;
+#       (6) the SHIPPED renderer, run as a subprocess on synthetic ServiceDefinitions, emits
+#           exactly the declared routes in order, emits NOTHING for a service declaring none or
+#           an empty list, and REFUSES all five breadth shapes.
+#
+# WHY ALSO A SCHEMA PATTERN, GIVEN THIS CHECK. The schema is the only surface that rejects a
+# too-broad regex WHERE A HUMAN TYPES IT — `ci/l0.sh` validates `catalog/*.yaml` against it on
+# every gate run, so the operator is told at the authoring surface rather than finding out from
+# a 401 on a page load. The renderer's own guard is there because `render-golden` runs even on a
+# gate invocation where the schema stage skipped. This check is there because neither of those
+# can see a renderer constant. Three surfaces, three different blind spots.
+#
+# WHY THIS IS A SIBLING AND NOT PART OF oauth2-proxy-cookie-names-distinct. That check's subject
+# is cookie IDENTITY: its inputs are names and domains, its failure mode is mutual session
+# eviction, its register is the shared proxy's claim on the default name. `--api-route` shares
+# none of those. Bundling them would make that check's name a lie and couple two unrelated
+# ratchets, so a future api-route decision could only be expressed by weakening a COOKIE check —
+# the same discipline that check's own header argues for, applied to this change.
+#
+# WHAT THIS DOES NOT ESTABLISH — read with the handoff entry, not instead of it:
+#   - That a 401 FIXES anything in a browser. No browser round-trip has been made, by anyone,
+#     for this flag. The 302/401 split, the dropped POST bodies and the orphaned SignalR poll
+#     are all read from access logs; the repair is reasoned from HTTP semantics.
+#   - That the regexes are RIGHT for these two apps beyond the measured paths. `^/api/` on
+#     radarr is justified by a measurement of radarr taken on its PRE-FIX pod by another
+#     engineer (decisions/2026-09-30-jackett-oauth2-proxy-session.md: 62 x /api/v3/command,
+#     46 x /api/v3/queue/details); the current pod has logged zero /api/ traffic.
+#   - Anything about sonarr, qbittorrent or pihole. Their logs were read (2026-10-01) and
+#     contain ZERO unauthenticated requests — no traffic, which is not the same as no need.
+#     Assertion (3) therefore proves the "declares none" PATH works; it does not claim those
+#     three apps are correctly configured.
+#   - That the LIVE pods carry what the tree says. git is not the cluster (CLAUDE.md §2b).
+#   - It does not read ../home-network-k3s. The spec is read-only (R-TEST-06).
+# =============================================================================
+apiVersion: chainsaw.kyverno.io/v1alpha1
+kind: Test
+metadata:
+  name: oauth2-proxy-api-routes-per-app
+  labels:
+    conformance.home.k3s/R-IDP-02: "true"
+  annotations:
+    conformance.home.k3s/requirements: "R-IDP-02"
+    conformance.home.k3s/layer: "L0"
+    # WHOLE-REPO SCOPE — the subject spans apps/wave-0, apps/wave-1, platform/crafty, every
+    # ServiceDefinition in catalog/, the renderer that generates five of the sidecars, and the
+    # ServiceDefinition schema. No single requirement path-maps to that set, and the renderer
+    # maps to no identity requirement at all, so a requirement-scoped selection would skip this
+    # check on precisely the change that can reintroduce the defect fleet-wide in one edit.
+    conformance.home.k3s/scope: "whole-repo"
+    conformance.home.k3s/safety-critical: "false"
+    conformance.home.k3s/executable: "true"
+    conformance.home.k3s/intent: >-
+      An unauthenticated XHR is answered 401, not 302, on exactly the paths each app was
+      MEASURED to serve as a machine API: every rendered --api-route is anchored at ^/ and
+      continues with literal path characters, every onboarded forward-auth service's rendered
+      routes equal its own auth.apiRoutes in order (so neither a hand-edited golden nor a
+      renderer constant can pass), an app declaring none renders none, the anchoring pattern is
+      the same in the schema and the renderer, and the shipped renderer is proven on synthetic
+      ServiceDefinitions in both directions — including refusing the five breadth shapes that
+      would answer a page load with 401.
+spec:
+  steps:
+    - name: assert-api-route-is-per-app-anchored-and-declared
+      description: >-
+        Inventory every oauth2-proxy sidecar's repeatable --api-route values, assert each is an
+        anchored path regex, cross-check each onboarded forward-auth service's rendered routes
+        against its own ServiceDefinition, require both the declares-routes and the
+        declares-none cases to be live in the tree, assert the renderer derives from
+        auth.apiRoutes and hardcodes no path regex, assert the schema and renderer share this
+        check's anchoring pattern, and prove the shipped renderer on synthetic
+        ServiceDefinitions in both directions.
+      try:
+        - script:
+            content: bash ./check.sh
@@ -0,0 +1,433 @@
+#!/usr/bin/env bash
+# =============================================================================
+# L0 STATIC CHECK — `--api-route` is PER APP, ANCHORED, and DECLARED.
+# See chainsaw-test.yaml's header for WHY this exists and what it measured. In one paragraph:
+#
+# oauth2-proxy answers an unauthenticated request with a 302 to Keycloak. A browser XHR follows
+# that redirect invisibly, the live SSO session satisfies it silently, and the call resolves 200
+# carrying the app's HTML where its JavaScript expected JSON — and a POST following a 302 is
+# re-issued as a GET with its BODY DROPPED. `--api-route=<regex>` makes matching paths answer
+# 401 instead. The flag is right for jackett (`/api/`) and radarr (`/signalr/`, `/api/`) and
+# WRONG for pihole, which serves no `/api/` at all — so it cannot be a renderer constant, and
+# the point of this check is that it never becomes one again.
+#
+# The breadth direction is the dangerous one and it is NOT symmetric with the bug. A missing
+# `--api-route` is a silent 302; a TOO BROAD one answers a plain PAGE LOAD with 401, so a
+# logged-out operator gets a bare error page and is never offered a sign-in — the service looks
+# permanently broken. That is a worse outcome than the defect, which is why the anchoring rule
+# is asserted here as well as in the schema.
+#
+# Exit 0 = all assertions pass. Exit 1 = a failure. Exit 3 (honest SKIP, which the runner turns
+# into rc 3 = UNWIRED, never green, because this test declares the whole-repo scope) only if the
+# tree contains no forward-auth oauth2-proxy sidecar at all — genuinely nothing to audit.
+# =============================================================================
+set -euo pipefail
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT="$(cd "${DIR}/../../../.." && pwd)"        # repo root
+PYTHON="${PYTHON:-python3}"
+
+# Honest SKIP (never green for a scope-declaring test) if the forward-auth pattern has left the
+# repo. Anchored on `image:` for the same reason as the sibling cookie check: a bare
+# `oauth2-proxy/oauth2-proxy:` also matches platform/oauth2-proxy/forward-auth-cec.yaml's Envoy
+# cluster reference (`serviceName: oauth2-proxy/oauth2-proxy:80`), which runs no proxy.
+if ! grep -rlE 'image:[[:space:]]*[^[:space:]]*oauth2-proxy' \
+       "${ROOT}/apps" "${ROOT}/platform" >/dev/null 2>&1; then
+  echo "check SKIPPED: no oauth2-proxy container image found under apps/ or platform/ — the forward-auth sidecar pattern is gone and there is nothing to audit" >&2
+  exit 3
+fi
+
+cd "${ROOT}"
+"${PYTHON}" - "${ROOT}" <<'PY'
+"""Static audit: --api-route is per-app, anchored, and declared in the ServiceDefinition.
+
+PyYAML (the same dependency as catalog/render.py and the sibling l0-static checks) rather than
+line-based parsing: a container's args live in a nested structure, and a line-based reader of
+one is precisely the approximate parser that silently stops matching.
+"""
+from __future__ import annotations
+
+import json
+import re
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+try:
+    import yaml
+except ImportError:                                  # pragma: no cover
+    print("check FAIL: PyYAML is required (same dependency as catalog/render.py)",
+          file=sys.stderr)
+    raise SystemExit(1)
+
+ROOT = Path(sys.argv[1])
+sys.dont_write_bytecode = True          # keep __pycache__ out of the tracked tree
+
+FAILURES: list[str] = []
+CHECKS = 0
+
+
+def check(cond, msg):
+    global CHECKS
+    CHECKS += 1
+    if not cond:
+        FAILURES.append(msg)
+
+
+# THE BREADTH RATCHET, stated once and asserted against three independent surfaces below (the
+# rendered tree, the authored catalog, the shipped renderer). Each entry must be anchored at
+# `^/` and continue with LITERAL path characters. This rejects:
+#   `^/`      and `^/.*`   — match every page load, so the UI answers 401 and never signs in
+#   `/api/`               — unanchored; oauth2-proxy would match it mid-path
+#   `.*`                  — everything
+#   `^/api/|^/`           — alternation smuggles breadth past a prefix-only reading
+# Alternation is REFUSED rather than supported because `--api-route` is repeatable: one entry
+# per route costs nothing and each entry stays readable. Kept byte-identical to
+# catalog/render.py's OAUTH2_API_ROUTE_PATTERN and to the schema's items.pattern; assertion (5)
+# below fails if they drift apart.
+ANCHORED = re.compile(r"^\^/[A-Za-z0-9._~/-]+(\.\*|\$)?$")
+SCHEMA_PATTERN = r"^\^/[A-Za-z0-9._~/-]+(\.\*|\$)?$"
+
+
+def read(p: Path) -> str:
+    try:
+        return p.read_text()
+    except (OSError, UnicodeDecodeError):
+        return ""
+
+
+def is_test_fixture(p: Path) -> bool:
+    """conformance/tests/**/{golden,fixtures}/** is assertion DATA, not deployed manifests."""
+    parts = p.parts
+    return "conformance" in parts and ("golden" in parts or "fixtures" in parts)
+
+
+def iter_containers(node):
+    """Yield every container mapping under any `containers`/`initContainers` list, at any depth.
+
+    Depth-agnostic on purpose: the same container shape appears under a Deployment, a DaemonSet,
+    a CronJob's jobTemplate or a bare Pod, and a reader that hardcoded one path would silently
+    stop seeing an instance the day one moved.
+    """
+    if isinstance(node, dict):
+        for key, val in node.items():
+            if key in ("containers", "initContainers") and isinstance(val, list):
+                for c in val:
+                    if isinstance(c, dict):
+                        yield c
+            else:
+                yield from iter_containers(val)
+    elif isinstance(node, list):
+        for item in node:
+            yield from iter_containers(item)
+
+
+def api_routes_of(args: list) -> list:
+    """Every `--api-route=<v>` / `--api-route <v>` value, in order. The flag is REPEATABLE, so
+    this must collect ALL of them — a single-value reader would audit the first and wave the
+    rest through, which is the shape of reader bug this file's header warns about."""
+    out = []
+    for i, a in enumerate(args):
+        if not isinstance(a, str):
+            continue
+        if a.startswith("--api-route="):
+            out.append(a.split("=", 1)[1])
+        elif a == "--api-route" and i + 1 < len(args) and isinstance(args[i + 1], str):
+            out.append(args[i + 1])
+    return out
+
+
+# ============================================================ (0) INVENTORY: THE RENDERED TREE
+# Every oauth2-proxy container this repo DEPLOYS, keyed by file + owner, with its api-routes.
+rendered: dict[str, list] = {}
+for p in sorted(ROOT.rglob("*.yaml")):
+    relparts = p.relative_to(ROOT).parent.parts
+    # NESTED WORKTREES: skip dot-DIRECTORIES (.git, .claude) while still matching dot-FILES such
+    # as .sops.yaml. .claude/worktrees/agent-*/ holds full repo copies for isolated agent runs,
+    # so a bare rglob audits up to 18 duplicates of every file — which both invents FAILs and,
+    # worse, can satisfy a "something declares X" check from a copy rather than from real source.
+    if any(part.startswith(".") for part in relparts) or is_test_fixture(p) or ".sops." in p.name:
+        continue
+    txt = read(p)
+    if "oauth2-proxy" not in txt:
+        continue
+    rel = str(p.relative_to(ROOT))
+    try:
+        docs = [d for d in yaml.safe_load_all(txt) if isinstance(d, dict)]
+    except yaml.YAMLError as exc:
+        check(False, f"{rel} is not parseable YAML ({exc}) — any oauth2-proxy instance in it "
+                     f"went unaudited, so a too-broad --api-route could be hiding there")
+        continue
+    for d in docs:
+        for c in iter_containers(d):
+            image = str(c.get("image") or "")
+            name = str(c.get("name") or "")
+            if "oauth2-proxy" not in image and name != "oauth2-proxy":
+                continue
+            owner = (d.get("metadata") or {}).get("name") or "?"
+            args = list(c.get("args") or []) + list(c.get("command") or [])
+            rendered[f"{rel}:{owner}/{name or 'oauth2-proxy'}"] = api_routes_of(args)
+
+check(len(rendered) >= 5,
+      f"only {len(rendered)} oauth2-proxy container(s) parsed out of the tree. This repo "
+      f"deploys more than that (five rendered sidecars plus crafty's hand-authored one), so the "
+      f"inventory has stopped seeing instances and every assertion below would pass vacuously "
+      f"while an unanchored --api-route went unaudited")
+
+# ============================================================ (1) EVERY ROUTE IS ANCHORED
+for key, routes in sorted(rendered.items()):
+    for r in routes:
+        check(bool(ANCHORED.match(r)),
+              f"{key} declares --api-route={r!r}, which is not an anchored path regex.\n"
+              f"       WHY THIS IS WORSE THAN THE BUG IT FIXES: --api-route makes matching paths "
+              f"answer 401 instead of 302ing to Keycloak. A regex that also matches a path a "
+              f"BROWSER NAVIGATES to — `^/`, `^/.*`, `.*`, or an alternation that reaches the "
+              f"UI root — answers a plain PAGE LOAD with 401. A logged-out operator then sees a "
+              f"bare error page and is NEVER OFFERED A SIGN-IN: the service looks permanently "
+              f"broken rather than merely unauthenticated, and no reload can recover it.\n"
+              f"       FIX: anchor it at '^/' followed by literal path characters "
+              f"(e.g. '^/api/', '^/signalr/', '^/api/v3/.*'), one entry per route — the flag is "
+              f"repeatable, so alternation is never needed. For a service rendered from catalog/, "
+              f"do NOT hand-edit the golden: declare auth.apiRoutes and re-render "
+              f"(ci/render-golden.sh's header has the canonical command)")
+
+# ============================================================ (2) DERIVED FROM THE CATALOG
+# The five rendered sidecars are OUTPUTS. Auditing only them would let someone "fix" a golden by
+# hand (reverted by the next render) or make the renderer emit a CONSTANT and regenerate every
+# golden to match — which render-golden cannot see, because render and golden would agree.
+# So: every onboarded forward-auth service's rendered route list must EQUAL its own
+# ServiceDefinition's auth.apiRoutes, in order.
+catalog_dir = ROOT / "catalog"
+declared: dict[str, list] = {}
+forward_auth: list[str] = []
+for sd_path in sorted(catalog_dir.glob("*.yaml")):
+    try:
+        sd = yaml.safe_load(read(sd_path))
+    except yaml.YAMLError as exc:
+        check(False, f"catalog/{sd_path.name} is not parseable YAML ({exc})")
+        continue
+    if not isinstance(sd, dict) or not sd.get("name"):
+        continue
+    auth = sd.get("auth") or {}
+    routes = auth.get("apiRoutes") or []
+    declared[sd["name"]] = list(routes)
+    if auth.get("mode") == "forward-auth" and (sd.get("ingressRoutes") or []):
+        forward_auth.append(sd["name"])
+    for r in routes:
+        check(bool(ANCHORED.match(str(r))),
+              f"catalog/{sd_path.name}: auth.apiRoutes entry {r!r} is not anchored at '^/' "
+              f"followed by literal path characters. See assertion (1) for why breadth here is "
+              f"the dangerous direction. The schema rejects this too (items.pattern); this "
+              f"assertion exists because the schema stage can be SKIPPED and this one cannot")
+    if routes:
+        check(auth.get("mode") == "forward-auth",
+              f"catalog/{sd_path.name} declares auth.apiRoutes with auth.mode="
+              f"{auth.get('mode')!r}. --api-route renders onto the per-app oauth2-proxy sidecar, "
+              f"and no sidecar is rendered for oidc/native/none — so this is a SILENT NO-OP: the "
+              f"operator believes the app's XHRs now answer 401 and nothing changed")
+
+matched = 0
+for name in sorted(forward_auth):
+    golden = next((p for p in sorted(ROOT.glob(f"apps/wave-*/{name}.yaml"))), None)
+    if golden is None:
+        continue                      # catalog service not onboarded to apps/ yet — no golden
+    key = next((k for k in rendered if k.startswith(str(golden.relative_to(ROOT)) + ":")), None)
+    check(key is not None,
+          f"apps/{golden.name} is the golden for forward-auth service {name!r} but no "
+          f"oauth2-proxy container was parsed out of it. Its --api-route configuration is "
+          f"UNAUDITED")
+    if key is None:
+        continue
+    matched += 1
+    check(rendered[key] == declared.get(name, []),
+          f"{name}: the rendered sidecar declares --api-route {rendered[key]!r} but "
+          f"catalog/{name}.yaml declares auth.apiRoutes {declared.get(name, [])!r}.\n"
+          f"       These must be IDENTICAL and in order. A mismatch means one of two things, "
+          f"both bad: a golden was hand-edited (the next re-render silently reverts it), or the "
+          f"renderer is emitting --api-route from somewhere other than this service's own "
+          f"ServiceDefinition — i.e. a constant or a per-service table. A constant is the defect "
+          f"this field exists to prevent: it applies ONE app's path shape to all five, and "
+          f"pihole serves no /api/ at all, so it would answer pihole's page loads with 401. "
+          f"render-golden cannot catch that, because render and golden would agree with each "
+          f"other while both being wrong")
+
+check(matched >= 4,
+      f"only {matched} onboarded forward-auth service(s) were cross-checked against their "
+      f"ServiceDefinition. This repo onboards five, so the catalog<->golden pairing has stopped "
+      f"resolving and assertion (2) is passing vacuously")
+
+# ============================================================ (3) BOTH CASES EXIST
+# Once the tree is consistent, (2) is satisfied by any pairing — including one where NOTHING
+# declares a route (vacuous) or one where EVERYTHING does (so the "renders nothing" path is
+# never exercised). Both halves are asserted to be live, because the per-app field only earns
+# its complexity if both outcomes actually occur in this tree.
+with_routes = [n for n in forward_auth if declared.get(n)]
+without_routes = [n for n in forward_auth if not declared.get(n)]
+check(bool(with_routes),
+      "no onboarded forward-auth service declares auth.apiRoutes, so assertion (2) cannot "
+      "distinguish a working renderer from one that emits the flag nowhere. If the flag is "
+      "genuinely not needed anywhere, delete the field rather than leaving an unexercised seam")
+check(bool(without_routes),
+      "EVERY onboarded forward-auth service declares auth.apiRoutes, so nothing in this tree "
+      "exercises the 'declares none => renders no flag' path. That path is the whole reason the "
+      "field is per-app: pihole serves no /api/, and a renderer-wide default would answer its "
+      "page loads with 401")
+for name in sorted(without_routes):
+    golden = next((p for p in sorted(ROOT.glob(f"apps/wave-*/{name}.yaml"))), None)
+    if golden is None:
+        continue
+    key = next((k for k in rendered if k.startswith(str(golden.relative_to(ROOT)) + ":")), None)
+    if key is None:
+        continue
+    check(rendered[key] == [],
+          f"{name} declares NO auth.apiRoutes but its rendered sidecar carries "
+          f"--api-route {rendered[key]!r}. An app that declares nothing must render NOTHING")
+
+# ============================================================ (4) THE RENDERER DERIVES
+render_py = ROOT / "catalog" / "render.py"
+check(render_py.is_file(),
+      "catalog/render.py is missing — the generator of five of these sidecars cannot be audited")
+if render_py.is_file():
+    src = read(render_py)
+    m = re.search(r"def _oauth2_sidecar_container\(.*?\n(?=\n*def |\n*# =)", src, re.S)
+    fn = m.group(0) if m else ""
+    check(bool(fn),
+          "could not locate _oauth2_sidecar_container() in catalog/render.py. It was renamed or "
+          "removed, so assertion (4) cannot run and the generator is unaudited")
+    if fn:
+        check("--api-route=" in fn,
+              "catalog/render.py's _oauth2_sidecar_container() emits no --api-route at all, so "
+              "every forward-auth app's XHRs are answered with a 302 the browser follows "
+              "invisibly — measured as 33x302-vs-2x401 on jackett and 232x302 on radarr, with "
+              "POST bodies dropped. auth.apiRoutes would then be a field with no consumer")
+        check(not re.search(r'--api-route=\^/[A-Za-z0-9]', fn),
+              "catalog/render.py hardcodes a literal path regex into --api-route. Whatever path "
+              "that is, EVERY forward-auth app now renders it: one app's path shape applied to "
+              "five. pihole serves no /api/, so a constant answers its page loads with 401 and "
+              "the service looks permanently broken to a logged-out operator. Render the routes "
+              "declared in each service's own auth.apiRoutes instead")
+        check("apiRoutes" in fn,
+              "catalog/render.py's _oauth2_sidecar_container() does not read auth.apiRoutes, so "
+              "the per-app declaration has no effect on what is rendered")
+
+# ============================================================ (5) ONE PATTERN, THREE SURFACES
+# The ratchet is written in three places (this check, the renderer, the schema) because each
+# guards a surface the others cannot. If they drift apart, the loosest one wins silently.
+if render_py.is_file():
+    # The renderer writes the pattern as a raw string literal, so the source text contains it
+    # verbatim. A substring test is deliberate: it also catches a rename of the constant.
+    check(SCHEMA_PATTERN in read(render_py),
+          f"catalog/render.py no longer contains the anchoring pattern {SCHEMA_PATTERN!r} that "
+          f"this check and the schema enforce. Three surfaces guard --api-route breadth (the "
+          f"authoring surface via the schema, the render via the renderer, the tree via this "
+          f"check); if they drift apart the loosest wins and the ratchet is decorative")
+
+# The schema's pattern is read from the PARSED JSON, not from the file text: in the file its
+# backslashes are JSON-escaped, so a substring test against the raw bytes would compare the
+# wrong string and pass or fail for the wrong reason.
+schema_file = ROOT / "schemas" / "servicedefinition.schema.json"
+schema_pat = None
+if schema_file.is_file():
+    try:
+        schema_pat = (((((json.loads(read(schema_file)).get("properties") or {})
+                         .get("auth") or {}).get("properties") or {})
+                       .get("apiRoutes") or {}).get("items") or {}).get("pattern")
+    except (ValueError, AttributeError) as exc:
+        check(False, f"schemas/servicedefinition.schema.json could not be read as JSON ({exc}) — "
+                     f"the schema stage that rejects a too-broad --api-route cannot be audited")
+check(schema_pat == SCHEMA_PATTERN,
+      f"schemas/servicedefinition.schema.json constrains auth.apiRoutes items with "
+      f"{schema_pat!r}, not {SCHEMA_PATTERN!r}. The schema is the only surface that rejects a "
+      f"too-broad regex WHERE A HUMAN TYPES IT (ci/l0.sh validates catalog/*.yaml on every gate "
+      f"run); if it is absent or looser than the renderer's, the operator finds out from a 401 "
+      f"on a page load")
+
+# ============================================================ (6) PROVE THE SHIPPED RENDERER
+# Once the tree is compliant, (1)-(3) have nothing left to catch and their passing is evidence
+# of nothing. This runs the SHIPPED renderer as a SUBPROCESS on synthetic ServiceDefinitions and
+# asserts the behaviour directly, in BOTH directions, so the guard stays falsifiable.
+def _sd(name: str, port: int, routes) -> str:
+    auth = {"mode": "forward-auth", "groups": ["platform-admins"]}
+    if routes is not None:
+        auth["apiRoutes"] = routes
+    return yaml.safe_dump({
+        "name": name,
+        "kind": "container",
+        "source": {"image": "example/app:1"},
+        "exposure": "internal",
+        "ingressRoutes": [{"host": f"{name}.home.keithjackson.dev", "port": port}],
+        "auth": auth,
+    }, sort_keys=False)
+
+
+def _render(name: str, port: int, routes):
+    with tempfile.TemporaryDirectory() as td:
+        f = Path(td) / f"{name}.yaml"
+        f.write_text(_sd(name, port, routes))
+        return subprocess.run([sys.executable, str(render_py), str(f)],
+                              capture_output=True, text=True)
+
+
+def _sidecar_args(rendered_yaml: str) -> list:
+    for doc in yaml.safe_load_all(rendered_yaml):
+        if not isinstance(doc, dict) or doc.get("kind") != "Deployment":
+            continue
+        for c in iter_containers(doc):
+            if c.get("name") == "oauth2-proxy":
+                return list(c.get("args") or [])
+    return []
+
+
+if render_py.is_file():
+    # (6a) DECLARED ROUTES RENDER, IN ORDER, AND NOTHING ELSE DOES.
+    proc = _render("apirtdecl", 8121, ["^/api/", "^/signalr/"])
+    check(proc.returncode == 0,
+          f"the shipped catalog/render.py failed on a synthetic forward-auth fixture declaring "
+          f"two apiRoutes: rc={proc.returncode} {proc.stderr.strip()[:300]}")
+    if proc.returncode == 0:
+        got = api_routes_of(_sidecar_args(proc.stdout))
+        check(got == ["^/api/", "^/signalr/"],
+              f"the shipped renderer turned auth.apiRoutes ['^/api/', '^/signalr/'] into "
+              f"--api-route {got!r}. It must emit one repeatable --api-route per entry, in "
+              f"declared order. Note the flag is SINGULAR: --api-routes does not exist in "
+              f"oauth2-proxy v7.8.1 and the proxy REFUSES TO START with it, so a plural would "
+              f"crashloop every sidecar it rendered")
+
+    # (6b) AN APP THAT DECLARES NOTHING RENDERS NOTHING. The pihole case, proven on the
+    #      generator rather than inferred from one committed golden.
+    for label, routes in (("absent", None), ("empty", [])):
+        proc = _render(f"apirtnone{label}", 8122, routes)
+        check(proc.returncode == 0,
+              f"the shipped renderer failed on a synthetic forward-auth fixture with apiRoutes "
+              f"{label}: rc={proc.returncode} {proc.stderr.strip()[:300]}")
+        if proc.returncode == 0:
+            got = api_routes_of(_sidecar_args(proc.stdout))
+            check(got == [],
+                  f"the shipped renderer emitted --api-route {got!r} for a synthetic service "
+                  f"whose auth.apiRoutes is {label}. A newly-onboarded forward-auth app must "
+                  f"render NO --api-route: pihole serves no /api/, so a default would answer "
+                  f"its page loads with 401 and the service would look permanently broken to a "
+                  f"logged-out operator — a worse failure than the 302 bug the flag fixes")
+
+    # (6c) THE RENDERER REFUSES BREADTH. Asserted on the generator because render-golden runs
+    #      on every gate invocation while the schema stage can be skipped.
+    for bad in ("^/", "^/.*", ".*", "/api/", "^/api/|^/"):
+        proc = _render("apirtbad", 8123, [bad])
+        check(proc.returncode != 0,
+              f"the shipped renderer ACCEPTED auth.apiRoutes {bad!r} and rendered a sidecar. "
+              f"That regex matches paths a browser NAVIGATES to, so a plain page load answers "
+              f"401 and the operator is never offered a sign-in. The renderer must refuse it, "
+              f"so a breadth mistake fails render-golden even on a run where the schema stage "
+              f"did not execute")
+
+# ============================================================ VERDICT
+inv = " | ".join(f"{k} -> {v or '(none)'}" for k, v in sorted(rendered.items()))
+print(f"oauth2-proxy sidecars audited ({len(rendered)}): {inv}")
+if FAILURES:
+    print(f"\nCHECK FAIL (oauth2-proxy-api-routes-per-app): {len(FAILURES)} failed, "
+          f"{CHECKS - len(FAILURES)} passed", file=sys.stderr)
+    for f in FAILURES:
+        print(f"  FAIL {f}", file=sys.stderr)
+    raise SystemExit(1)
+print(f"check OK: oauth2-proxy-api-routes-per-app — {CHECKS} assertions (R-IDP-02)")
+PY
@@ -0,0 +1,152 @@
+# =============================================================================
+# L0 STATIC — no two oauth2-proxy instances can share a browser cookie
+# Catalog: docs/14 L0 row ("Pre-apply, in CI") — runs WITHOUT a cluster.
+#
+# THE DEFECT THIS CLOSES. Every oauth2-proxy on this platform is an INDEPENDENT instance with
+# its OWN `cookie-secret`: five per-app reverse-proxy sidecars (jackett, pihole, qbittorrent,
+# radarr, sonarr), crafty's, and the SHARED forward-auth proxy. Measured on prod 2026-09-30:
+# seven `*-oauth2-proxy` Secrets, SEVEN DISTINCT cookie-secret values — so no instance can
+# decrypt any other's cookie.
+#
+# They were all emitting oauth2-proxy's DEFAULT cookie names (`_oauth2_proxy` and
+# `_oauth2_proxy_csrf`), scoped to `.${BASE_DOMAIN}`. A cookie is keyed by (name, domain, path),
+# so the browser kept ONE pair for ALL of them. Whichever proxy authenticated last OVERWROTE
+# every other's session; the loser logged `No valid authentication in request`, bounced to
+# Keycloak, and — the SSO session still being live — was redirected straight back and overwrote
+# the cookie again. Six live instances, one cookie, mutual eviction.
+#
+# WHY IT IS THE WORST KIND OF BUG. The symptom is "I signed into Keycloak but I'm still not
+# signed in", which points the investigator at Keycloak, where nothing is wrong. It is
+# INTERMITTENT — it only appears once a browser has visited a SECOND gated app, so the obvious
+# single-app reproduction succeeds — and it is SILENT: every Deployment stays Ready, every
+# Kustomization stays Ready, the proxy's own log line names an authentication failure rather
+# than a collision. It went unnoticed across five services for the whole life of the pattern,
+# and was found only as a side effect of investigating something else (2026-09-29, crafty).
+#
+# WHY A COMMIT IS NOT ENOUGH, AND THIS IS A GATE STAGE. The collision is a property of the
+# PATTERN, not of any one app: the next forward-auth service onboarded inherits the defect by
+# doing nothing at all, because a missing flag IS the bug. Nothing about its absence is loud.
+# This is the same shape as the twelve inert ServiceMonitors in this tree — one omission,
+# copied, invisible for weeks — and the same remedy: make the omission fail the merge.
+#
+# GIVEN every oauth2-proxy instance this repo authors — a container in any manifest (the
+#       rendered sidecars in apps/, crafty's hand-authored one in platform/) and the shared
+#       proxy's HelmRelease values,
+# WHEN  each instance's EFFECTIVE cookie name is resolved (its declared `--cookie-name` /
+#       `OAUTH2_PROXY_COOKIE_NAME`, or oauth2-proxy's built-in default `_oauth2_proxy` when it
+#       declares none — an omission is a NAME, not an absence, and that is the whole bug),
+# THEN  (1) no two instances share an effective cookie name;
+#       (2) no instance's effective name is another's name + `_csrf` — oauth2-proxy derives the
+#           CSRF cookie that way, so `_oauth2_proxy` and `_oauth2_proxy_csrf` as two SESSION
+#           names would collide on the second cookie while looking distinct on the first;
+#       (3) the DEFAULT name has exactly ONE owner and it is the instance declared in
+#           SHARED_DEFAULT_COOKIE_OWNER. This is the deliberate ASYMMETRY. The shared
+#           forward-auth proxy MUST keep `_oauth2_proxy`: it is the single session every
+#           OTHER forward-auth-gated surface on the platform rides on, so renaming it signs
+#           every one of them out at once. The rule is therefore NOT "every instance must
+#           declare a name" — it is "the default is a NAME LIKE ANY OTHER, held by exactly one
+#           REGISTERED owner". That formulation needs no exemption from (1), and it closes the
+#           hole an exemption would leave: if the shared proxy is ever removed, an app silently
+#           inheriting the default would still FAIL here instead of quietly becoming its heir;
+#       (4) every instance that is NOT that owner is HOST-SCOPED: its `--cookie-domain` equals
+#           the host of its own `--redirect-url`. See the argument below;
+#       (5) catalog/render.py DERIVES both flags from the ServiceDefinition (the service name
+#           and the route host) and does not carry a per-service table. A list is how twelve
+#           ServiceMonitors acquired one identical wrong port name in a single edit, and a list
+#           leaves the NEXT app defaulting into the collision;
+#       (6) the SHIPPED renderer is proven END-TO-END on synthetic ServiceDefinitions, run as a
+#           subprocess: two forward-auth fixtures must yield two different, non-default cookie
+#           names, each scoped to its own host. Without this, once the tree is compliant (1)-(4)
+#           are satisfied by a tree with nothing left to catch, and a guard whose only evidence
+#           is "nothing failed" cannot be told apart from one that stopped working;
+#   AND (7) the check is LOAD-BEARING: this Test DECLARES `conformance.home.k3s/scope:
+#           whole-repo`, so ci/l0.sh Stage 3b's `--include-scope whole-repo` unions it into
+#           EVERY gate run, and ci/gate.sh blocks the merge when the l0 stage fails. It
+#           deliberately adds NO stage to ci/l0.sh — tests/l0-static/whole-repo-invariants-wired
+#           fails the gate if anyone adds one.
+#
+# WHY SCOPE, NOT REQUIREMENTS: ci/touched-requirements.sh maps `apps/*` to R-IDP-02/R-IDP-06
+# today, but the instances live in THREE unrelated subtrees (apps/, platform/crafty/,
+# platform/oauth2-proxy/) and the file most likely to reintroduce the defect is
+# catalog/render.py, which maps to R-GITOPS-02 and to no identity requirement at all. A
+# requirement-scoped selection would filter this check out of exactly the change that breaks it.
+#
+# ON ASSERTION (4), HOST-SCOPED COOKIE-DOMAIN — the argument, and the decision.
+#   AGAINST asserting it (overreach): distinct NAMES already close the eviction bug completely.
+#   Two cookies with different names do not overwrite each other however wide their domains are,
+#   so (4) fixes nothing that (1) has not already fixed. `--cookie-domain` is also a legitimate
+#   knob — a future proxy fronting two hostnames of one app would need a wider value — and a
+#   check that fails correct configurations is a check someone eventually disables.
+#   FOR asserting it (and this wins): a zone-wide cookie is TRANSMITTED BY THE BROWSER TO EVERY
+#   HOST IN THE ZONE. qBittorrent's session token is therefore sent to pihole, to radarr, to
+#   sonarr, and to every other name under `.${BASE_DOMAIN}` — including hosts behind no proxy at
+#   all. That is a session-credential disclosure across trust boundaries that no one needs and
+#   nobody asked for; the sidecar's session is only ever meaningful to the one app it fronts.
+#   And it is what makes a FUTURE name collision catastrophic rather than contained: with
+#   host-scoped cookies, two instances that accidentally share a name cannot even see each
+#   other's cookie, so the failure this whole test exists to prevent degrades from platform-wide
+#   to nothing. Defence in depth on the exact axis that already failed once.
+#   DECISION: assert it, for every instance except the registered shared proxy — whose whole
+#   job IS cross-host forward-auth, and which is exempted by name and with a reason, not by
+#   loosening the rule. `--whitelist-domain` is deliberately NOT touched: it governs permitted
+#   post-login REDIRECT targets, not cookie scope, and narrowing it breaks the sign-in bounce.
+#
+# STATUS: EXECUTABLE. `check.sh` is an offline assertion script (PyYAML, as catalog/render.py
+# and the sibling l0-static checks use); conformance/run.sh --layer l0 and ci/l0.sh run it
+# directly, and this Chainsaw Test wraps the SAME script.
+#
+# NOT proven (honest gaps; this is STATIC, with no cluster and no browser):
+#   - That anyone can actually sign in. No browser round-trip is performed, and none can be
+#     from L0. The 2026-09-29 crafty entry recorded the same gap and it is still open: the
+#     symptom has never been REPRODUCED in a browser, so this is a latent-bug fix reasoned from
+#     measured configuration, not a fix watched to work.
+#   - That the LIVE pods carry what the tree says. git is not the cluster (CLAUDE.md §2b). The
+#     live flags were measured separately, once, by hand; nothing here re-measures them, and a
+#     hand-applied drift would not be visible to this check.
+#   - That the cookie-secrets really differ. That is a Secret comparison and belongs in-cluster;
+#     it was measured out of band (seven distinct values, compared without printing any of
+#     them). If they were ever made IDENTICAL the collision would become harmless — but sharing
+#     one session-encryption key across trust zones is a worse idea than the bug it would mask,
+#     so this check does not offer it as an alternative and does not look for it.
+#   - Anything about the CONTENT of a session, group authorization, or the CSRF cookie's own
+#     behaviour. Names and domains only.
+#   - It does not read ../home-network-k3s. The spec is read-only (R-TEST-06).
+# =============================================================================
+apiVersion: chainsaw.kyverno.io/v1alpha1
+kind: Test
+metadata:
+  name: oauth2-proxy-cookie-names-distinct
+  labels:
+    conformance.home.k3s/R-IDP-02: "true"
+    conformance.home.k3s/R-IDP-06: "true"
+  annotations:
+    conformance.home.k3s/requirements: "R-IDP-02, R-IDP-06"
+    conformance.home.k3s/layer: "L0"
+    # WHOLE-REPO SCOPE — the instances span apps/wave-0, apps/wave-1, platform/crafty and
+    # platform/oauth2-proxy, and the generator that emits five of them is catalog/render.py.
+    # No single requirement path-maps to that set, and the renderer maps to no identity
+    # requirement at all, so a requirement-scoped selection would skip this check on precisely
+    # the change that can reintroduce the defect fleet-wide in one edit.
+    conformance.home.k3s/scope: "whole-repo"
+    conformance.home.k3s/safety-critical: "false"
+    conformance.home.k3s/executable: "true"
+    conformance.home.k3s/intent: >-
+      No two oauth2-proxy instances can share a browser cookie: every instance's effective
+      cookie name is unique, the built-in default belongs to exactly one registered owner (the
+      shared forward-auth proxy, which cannot be renamed without signing every gated app out),
+      every other instance is host-scoped, and catalog/render.py derives both flags from the
+      ServiceDefinition rather than from a per-service list — proven end-to-end on synthetic
+      fixtures so the guard stays falsifiable once the tree is clean.
+spec:
+  steps:
+    - name: assert-no-two-oauth2-proxy-instances-share-a-cookie
+      description: >-
+        Inventory every oauth2-proxy container and HelmRelease in the tree, resolve each one's
+        effective cookie name (declared, or oauth2-proxy's default when omitted), assert
+        pairwise distinctness including the session/CSRF derivation, ratchet the single
+        registered owner of the default name, assert host-scoped cookie-domain on every other
+        instance, assert catalog/render.py derives rather than lists, and prove the shipped
+        renderer on two synthetic forward-auth ServiceDefinitions.
+      try:
+        - script:
+            content: bash ./check.sh
@@ -0,0 +1,498 @@
+#!/usr/bin/env bash
+# =============================================================================
+# L0 STATIC CHECK — no two oauth2-proxy instances can share a browser cookie.
+# See chainsaw-test.yaml's header for WHY this exists, what it measured, and the argument
+# for assertion (4). In one paragraph:
+#
+# Every oauth2-proxy here holds its OWN cookie-secret (seven distinct values on prod,
+# 2026-09-30), so two instances sharing a cookie NAME on a shared DOMAIN cannot read each
+# other's session — they OVERWRITE it. The loser bounces to Keycloak, is redirected straight
+# back by the live SSO session, and overwrites the winner in turn. The symptom is "I signed
+# into Keycloak but I'm still not signed in", it only shows up after a second gated app is
+# visited, and nothing in the cluster reports an error. A missing --cookie-name IS the bug,
+# so the next forward-auth app inherits it by doing nothing.
+#
+# Exit 0 = all assertions pass. Exit 1 = a failure. Exit 3 (honest SKIP, which the runner
+# turns into rc 3 = UNWIRED, never green, because this test declares the whole-repo scope)
+# only if the tree contains no oauth2-proxy instance at all — genuinely nothing to audit.
+# =============================================================================
+set -euo pipefail
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT="$(cd "${DIR}/../../../.." && pwd)"        # repo root
+PYTHON="${PYTHON:-python3}"
+
+# Honest SKIP (never green for a scope-declaring test) if there is nothing to audit at all.
+# Matches the IMAGE every instance runs and the CHART the shared one installs; if NEITHER
+# appears anywhere, the forward-auth pattern has left this repo and the check has no subject.
+#
+# ANCHORED ON `image:` DELIBERATELY. A bare `oauth2-proxy/oauth2-proxy:` also matches
+# platform/oauth2-proxy/forward-auth-cec.yaml's Envoy cluster reference
+# (`serviceName: oauth2-proxy/oauth2-proxy:80` — namespace/service:port, not a registry path).
+# That file configures the ext_authz listener that CALLS the proxy; it runs no proxy and holds
+# no cookie flags, so counting it made the grep/parse cross-check below report a phantom
+# unaudited instance.
+if ! grep -rlE 'image:[[:space:]]*[^[:space:]]*oauth2-proxy|chart:[[:space:]]*oauth2-proxy' \
+       "${ROOT}/apps" "${ROOT}/platform" "${ROOT}/catalog" >/dev/null 2>&1; then
+  echo "check SKIPPED: no oauth2-proxy image or chart found under apps/, platform/ or catalog/ — nothing to audit" >&2
+  exit 3
+fi
+
+cd "${ROOT}"
+"${PYTHON}" - "${ROOT}" <<'PY'
+"""Static audit: no two oauth2-proxy instances can share a browser cookie.
+
+PyYAML (the same dependency as catalog/render.py and the sibling l0-static checks) rather
+than line-based parsing: a container's args live in a nested structure and a HelmRelease's
+flags live in a values map, and a line-based reader of either is precisely the approximate
+parser that silently stops matching — the failure class this check exists to prevent.
+"""
+from __future__ import annotations
+
+import re
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+try:
+    import yaml
+except ImportError:                                  # pragma: no cover
+    print("check FAIL: PyYAML is required (same dependency as catalog/render.py)",
+          file=sys.stderr)
+    raise SystemExit(1)
+
+ROOT = Path(sys.argv[1])
+sys.dont_write_bytecode = True          # keep __pycache__ out of the tracked tree
+
+FAILURES: list[str] = []
+CHECKS = 0
+
+
+def check(cond, msg):
+    global CHECKS
+    CHECKS += 1
+    if not cond:
+        FAILURES.append(msg)
+
+
+def read(p: Path) -> str:
+    try:
+        return p.read_text()
+    except (OSError, UnicodeDecodeError):
+        return ""
+
+
+# oauth2-proxy's OWN built-in defaults (pkg/apis/options/cookie.go): the session cookie is
+# `_oauth2_proxy` and the CSRF cookie is that name with `_csrf` appended. An instance that
+# declares nothing is NOT unconfigured — it has claimed this exact name, which is why the
+# inventory below treats the default as a name like any other rather than as an absence.
+DEFAULT_COOKIE_NAME = "_oauth2_proxy"
+CSRF_SUFFIX = "_csrf"
+
+# =============================================================================================
+# REGISTER — SHARED_DEFAULT_COOKIE_OWNER. Exactly one entry, and it must match a real instance.
+#
+# The one instance permitted to hold DEFAULT_COOKIE_NAME, and the one permitted a zone-wide
+# cookie-domain. Both exemptions belong to the SAME object for the SAME reason, so they are one
+# register rather than two: the shared forward-auth proxy is not a per-app sidecar, it is the
+# single session that every forward-auth-gated surface on the platform rides on. Renaming its
+# cookie signs all of them out simultaneously, and host-scoping it would break the cross-host
+# forward-auth that is its entire job.
+#
+# Shrink-only, and exact in BOTH directions: an entry that matches no instance FAILS as stale
+# (the proxy was renamed or removed and the exemption is now a hole), and an instance holding
+# the default WITHOUT an entry FAILS as an unregistered heir. Adding an entry is a review event
+# on a sensitive path.
+#
+# Keyed by "<relpath>:<instance id>".
+# =============================================================================================
+SHARED_DEFAULT_COOKIE_OWNER: dict[str, str] = {
+    "platform/oauth2-proxy/helmrelease.yaml:oauth2-proxy":
+        "The SHARED forward-auth proxy (ADR-0014, platform/oauth2-proxy/README.md). It is the "
+        "single authenticator behind every forward-auth-gated surface that is NOT fronted by "
+        "its own sidecar, so its cookie is the platform session: renaming it signs every one "
+        "of them out at once, and that is why the 2026-09-29 crafty fix changed only the "
+        "sidecar side. Its cookie-domain is zone-wide for the same structural reason — "
+        "cross-host forward-auth is what it is for.",
+}
+
+
+def is_test_fixture(p: Path) -> bool:
+    """conformance/tests/**/{golden,fixtures}/** is assertion DATA, not deployed manifests.
+
+    Auditing a golden would be auditing a copy of something already audited through apps/.
+    Same exclusion the sibling l0-static checks make, and the same "reference, not
+    declaration" distinction ci/touched-requirements.sh draws for fixture paths.
+    """
+    parts = p.parts
+    return "conformance" in parts and ("golden" in parts or "fixtures" in parts)
+
+
+def iter_containers(node):
+    """Yield every container mapping under any `containers`/`initContainers` list, at any depth.
+
+    Depth-agnostic on purpose: a sidecar sits under spec.template.spec.containers in a
+    Deployment, but the same container shape appears under a CronJob's jobTemplate, a
+    DaemonSet, or a Pod, and a reader that hardcoded one path would silently stop seeing an
+    instance the day one moved.
+    """
+    if isinstance(node, dict):
+        for key, val in node.items():
+            if key in ("containers", "initContainers") and isinstance(val, list):
+                for c in val:
+                    if isinstance(c, dict):
+                        yield c
+            else:
+                yield from iter_containers(val)
+    elif isinstance(node, list):
+        for item in node:
+            yield from iter_containers(item)
+
+
+def flag_from_args(args: list, flag: str):
+    """Value of `--flag=value` or `--flag value` in an argv list, else None."""
+    pref = "--" + flag
+    for i, a in enumerate(args):
+        if not isinstance(a, str):
+            continue
+        if a.startswith(pref + "="):
+            return a.split("=", 1)[1]
+        if a == pref and i + 1 < len(args) and isinstance(args[i + 1], str):
+            return args[i + 1]
+    return None
+
+
+def host_of(url):
+    """Host part of an https://host/path URL, else None."""
+    if not isinstance(url, str):
+        return None
+    m = re.match(r"^https?://([^/]+)", url)
+    return m.group(1) if m else None
+
+
+# ============================================================ (0) INVENTORY
+# instances: list of dicts {rel, ident, kind, declared_name, cookie_domain, redirect_host}
+instances: list[dict] = []
+unparseable: list[str] = []
+# Files the grep gate says contain an instance, so the parse can be cross-checked against it.
+grep_files: set[str] = set()
+
+for p in sorted(ROOT.rglob("*.yaml")):
+    relparts = p.relative_to(ROOT).parent.parts
+    # NESTED WORKTREES: skip dot-DIRECTORIES (.git, .claude) while still matching dot-FILES
+    # such as .sops.yaml. .claude/worktrees/agent-*/ holds full repo copies for isolated agent
+    # runs, so a bare rglob audited up to 18 duplicates of every file — which both invents
+    # FAILs and, worse, can satisfy a "something declares X" check from a copy rather than from
+    # real source. Same fix as egress-paths-selected-by-rules. .claude/ is gitignored.
+    if any(part.startswith(".") for part in relparts) or is_test_fixture(p) or ".sops." in p.name:
+        continue
+    txt = read(p)
+    if not txt:
+        continue
+    rel = str(p.relative_to(ROOT))
+    # Same anchoring as check.sh's skip gate: `image:` (or a chart name), never a bare path —
+    # see the comment there for the Envoy `serviceName: oauth2-proxy/oauth2-proxy:80` near-miss.
+    if re.search(r"image:\s*\S*oauth2-proxy|chart:\s*oauth2-proxy", txt):
+        grep_files.add(rel)
+    elif "oauth2-proxy" not in txt:
+        continue
+    try:
+        docs = [d for d in yaml.safe_load_all(txt) if isinstance(d, dict)]
+    except yaml.YAMLError as exc:
+        unparseable.append(f"{rel} ({exc})")
+        continue
+    for d in docs:
+        # --- shape 1: a container running the oauth2-proxy image -------------------------
+        for c in iter_containers(d):
+            image = str(c.get("image") or "")
+            name = str(c.get("name") or "")
+            if "oauth2-proxy" not in image and name != "oauth2-proxy":
+                continue
+            args = [a for a in (list(c.get("args") or []) + list(c.get("command") or []))]
+            env = {e.get("name"): e.get("value")
+                   for e in (c.get("env") or []) if isinstance(e, dict) and "value" in e}
+            owner = (d.get("metadata") or {}).get("name") or "?"
+            instances.append({
+                "rel": rel,
+                "ident": f"{owner}/{name or 'oauth2-proxy'}",
+                "kind": "container",
+                "declared_name": flag_from_args(args, "cookie-name")
+                                 or env.get("OAUTH2_PROXY_COOKIE_NAME"),
+                "cookie_domain": flag_from_args(args, "cookie-domain")
+                                 or env.get("OAUTH2_PROXY_COOKIE_DOMAIN"),
+                "redirect_host": host_of(flag_from_args(args, "redirect-url")
+                                         or env.get("OAUTH2_PROXY_REDIRECT_URL")),
+            })
+        # --- shape 2: the shared proxy's HelmRelease values ------------------------------
+        if d.get("kind") != "HelmRelease":
+            continue
+        spec = d.get("spec") or {}
+        chart = (((spec.get("chart") or {}).get("spec") or {}).get("chart") or "")
+        if "oauth2-proxy" not in str(chart):
+            continue
+        values = spec.get("values") or {}
+        extra = values.get("extraArgs") or {}
+        if isinstance(extra, list):          # the chart also accepts a bare argv list
+            extra = {a.lstrip("-").split("=", 1)[0]: a.split("=", 1)[1]
+                     for a in extra if isinstance(a, str) and "=" in a}
+        cfg = values.get("config") or {}
+        get = lambda k: extra.get(k, cfg.get(k))     # noqa: E731 - two-source lookup, used 3x
+        instances.append({
+            "rel": rel,
+            "ident": str((d.get("metadata") or {}).get("name") or "oauth2-proxy"),
+            "kind": "helmrelease",
+            "declared_name": get("cookie-name"),
+            "cookie_domain": get("cookie-domain"),
+            "redirect_host": host_of(get("redirect-url")),
+        })
+
+for u in unparseable:
+    check(False, f"{u} is not parseable YAML — every oauth2-proxy instance in it went unaudited")
+
+for ins in instances:
+    ins["key"] = f"{ins['rel']}:{ins['ident']}"
+    # THE CORE MODELLING DECISION: an omitted --cookie-name is not "unset", it is a CLAIM on
+    # oauth2-proxy's built-in default. Resolving it here is what lets assertion (1) below be a
+    # plain distinctness test with no special case, and it is exactly the step whose absence
+    # made the defect invisible to every reviewer who read these manifests.
+    ins["effective_name"] = ins["declared_name"] or DEFAULT_COOKIE_NAME
+
+# Anti-vacuous. The floor is deliberately low (a real removal should not turn this check red),
+# so the load-bearing guard is the grep/parse CROSS-CHECK beneath it: every file the cheap
+# text scan says holds an instance must yield at least one PARSED instance. A parser that
+# quietly stopped matching would keep the floor and fail this.
+check(len(instances) >= 5,
+      f"only {len(instances)} oauth2-proxy instance(s) parsed out of the tree. This repo "
+      f"authors more than that (five rendered sidecars, crafty's, and the shared proxy), so "
+      f"the inventory has stopped seeing instances and every assertion below would pass "
+      f"vacuously while a collision went unaudited")
+
+parsed_files = {ins["rel"] for ins in instances}
+for rel in sorted(grep_files - parsed_files):
+    check(False,
+          f"{rel} contains an oauth2-proxy image or chart reference but NO instance was parsed "
+          f"out of it. Either it uses a shape iter_containers()/the HelmRelease reader does not "
+          f"recognise, or the file stopped parsing. Its cookie configuration is UNAUDITED, which "
+          f"is the same state the whole fleet was in before this check existed")
+
+# ============================================================ (1) DISTINCT EFFECTIVE NAMES
+by_name: dict[str, list[dict]] = {}
+for ins in instances:
+    by_name.setdefault(ins["effective_name"], []).append(ins)
+
+for nm, group in sorted(by_name.items()):
+    if len(group) < 2:
+        continue
+    who = ", ".join(sorted(g["key"] for g in group))
+    default_note = (
+        " Every one of them OMITS --cookie-name, so they are all claiming oauth2-proxy's "
+        "built-in default. Omitting the flag is not neutral: it IS this collision."
+        if nm == DEFAULT_COOKIE_NAME else "")
+    check(False,
+          f"{len(group)} oauth2-proxy instances share the cookie name '{nm}': {who}.{default_note}\n"
+          f"       WHY THIS IS A BUG AND NOT A STYLE RULE: these are INDEPENDENT instances with "
+          f"DIFFERENT cookie-secrets (measured on prod 2026-09-30: seven distinct values), so "
+          f"neither can DECRYPT the other's cookie — it can only OVERWRITE it. A browser keys a "
+          f"cookie by (name, domain, path), so signing into one app destroys the session of every "
+          f"other app sharing that key. The loser logs 'No valid authentication in request', "
+          f"bounces to Keycloak, is redirected straight back because the SSO session is still "
+          f"live, and overwrites the winner in turn. The user sees 'I signed into Keycloak but "
+          f"I'm still not signed in' and goes looking in Keycloak, where nothing is wrong.\n"
+          f"       FIX: give each per-app instance its own --cookie-name. For a service rendered "
+          f"from catalog/, do NOT hand-edit the golden — catalog/render.py derives the flag from "
+          f"the service name; re-render instead (ci/render-golden.sh's header has the canonical "
+          f"command). Exactly one instance may hold '{DEFAULT_COOKIE_NAME}', and it must be the "
+          f"one registered in SHARED_DEFAULT_COOKIE_OWNER.")
+
+# ============================================================ (2) SESSION vs CSRF DERIVATION
+# oauth2-proxy names the CSRF cookie <cookie-name>_csrf. So a SESSION name that happens to equal
+# another instance's name + "_csrf" collides on the CSRF cookie while assertion (1) sees two
+# perfectly distinct strings. This is the near-miss the obvious fix invites: naming the shared
+# proxy's neighbour `_oauth2_proxy_csrf` reads as "clearly different" and is not.
+effective = {ins["effective_name"] for ins in instances}
+for ins in sorted(instances, key=lambda i: i["key"]):
+    base = ins["effective_name"]
+    if not base.endswith(CSRF_SUFFIX):
+        continue
+    victim = base[: -len(CSRF_SUFFIX)]
+    check(victim not in effective,
+          f"{ins['key']} uses the SESSION cookie name '{base}', which is exactly the CSRF cookie "
+          f"oauth2-proxy derives for the instance named '{victim}' (it appends '{CSRF_SUFFIX}'). "
+          f"The two session names look distinct and the CSRF cookies collide, so sign-in breaks "
+          f"with a state-mismatch instead of a session loss — the same bug wearing a different "
+          f"error message. Pick a name that is not another instance's name + '{CSRF_SUFFIX}'")
+
+# ============================================================ (3) THE DEFAULT HAS ONE OWNER
+default_holders = sorted(ins["key"] for ins in instances
+                         if ins["effective_name"] == DEFAULT_COOKIE_NAME)
+registered = set(SHARED_DEFAULT_COOKIE_OWNER)
+
+for key in sorted(registered - set(ins["key"] for ins in instances)):
+    check(False,
+          f"SHARED_DEFAULT_COOKIE_OWNER registers '{key}', but no such oauth2-proxy instance "
+          f"exists in the tree. The register is SHRINK-ONLY and exact: a stale entry is a "
+          f"standing exemption for an object that is gone, so the next instance to inherit the "
+          f"default name would be silently excused. Remove the entry, or restore the instance")
+
+for key in default_holders:
+    check(key in registered,
+          f"{key} holds oauth2-proxy's DEFAULT cookie name '{DEFAULT_COOKIE_NAME}' but is NOT "
+          f"registered in SHARED_DEFAULT_COOKIE_OWNER. Exactly ONE instance may hold the default, "
+          f"and it is the SHARED forward-auth proxy — whose cookie is the platform-wide session "
+          f"every forward-auth-gated surface rides on, so renaming IT signs them all out at once. "
+          f"Every other instance must declare its own --cookie-name. If this instance really is "
+          f"the new shared proxy, register it here with the reason; if it is a per-app sidecar, "
+          f"give it a unique name")
+
+check(len(default_holders) <= 1,
+      f"{len(default_holders)} instances hold the default cookie name: {', '.join(default_holders)}. "
+      f"The default is a NAME, not an absence — it can have exactly one owner")
+
+# ============================================================ (4) HOST-SCOPED COOKIE DOMAIN
+# See the chainsaw-test header for the argument (distinct names alone already fix the eviction;
+# host-scoping is asserted anyway because a zone-wide cookie is TRANSMITTED to every host in the
+# zone, and because it makes a future name collision contained instead of platform-wide).
+for ins in sorted(instances, key=lambda i: i["key"]):
+    if ins["key"] in registered:
+        continue                          # the shared proxy: zone-wide by design, registered
+    dom, rhost = ins["cookie_domain"], ins["redirect_host"]
+    check(dom is not None,
+          f"{ins['key']} declares no --cookie-domain. oauth2-proxy then scopes the cookie to the "
+          f"request host, which happens to be correct — but it is correct BY ACCIDENT and says "
+          f"nothing to the next reader. Declare it explicitly as this app's own host")
+    if dom is None:
+        continue
+    check(not dom.startswith("."),
+          f"{ins['key']} sets --cookie-domain={dom}. A leading dot makes the cookie ZONE-WIDE: "
+          f"the browser sends this app's session token to every other host under that domain, "
+          f"including hosts behind no proxy at all, and it is what turns any future cookie-name "
+          f"collision into a platform-wide outage instead of a contained one. Scope it to this "
+          f"app's own host. Only the registered shared forward-auth proxy may be zone-wide, "
+          f"because cross-host forward-auth is its job")
+    if rhost:
+        check(dom == rhost,
+              f"{ins['key']} sets --cookie-domain={dom} but its --redirect-url host is {rhost}. "
+              f"The cookie is then set for a domain the sign-in flow never returns to (or for a "
+              f"wider one than this app owns), so the session either does not stick or leaks "
+              f"beyond the app. Both flags describe the SAME app: make them agree")
+
+# ============================================================ (5) THE RENDERER DERIVES
+# Five of these instances are GENERATED. Auditing only the goldens would let someone "fix" the
+# tree by hand and leave the generator emitting the defect for the next service onboarded — and
+# the next re-render would then silently revert the hand fix.
+render_py = ROOT / "catalog" / "render.py"
+check(render_py.is_file(),
+      "catalog/render.py is missing — the generator of five of these instances cannot be audited")
+if render_py.is_file():
+    src = read(render_py)
+    m = re.search(r"def _oauth2_sidecar_container\(.*?\n(?=\n*def |\n*# =)", src, re.S)
+    fn = m.group(0) if m else ""
+    check(bool(fn),
+          "could not locate _oauth2_sidecar_container() in catalog/render.py. It was renamed or "
+          "removed; assertions (5) below cannot run, and the generator is unaudited")
+    if fn:
+        cookie_name_lines = [ln for ln in fn.splitlines() if "--cookie-name" in ln]
+        check(bool(cookie_name_lines),
+              "catalog/render.py's _oauth2_sidecar_container() emits NO --cookie-name at all, so "
+              "every service it renders claims oauth2-proxy's default and collides with every "
+              "other. This is the defect itself, back in the one place that reintroduces it "
+              "fleet-wide in a single edit")
+        check(any("s['name']" in ln or 's["name"]' in ln for ln in cookie_name_lines),
+              "catalog/render.py emits --cookie-name without DERIVING it from the "
+              "ServiceDefinition's own name. A constant collides with every sibling by "
+              "construction; a per-service table leaves the NEXT app onboarded defaulting into "
+              "the collision, which is how twelve ServiceMonitors acquired one identical wrong "
+              "port name in a single edit. Derive it from s['name']")
+        cookie_dom_lines = [ln for ln in fn.splitlines() if "--cookie-domain" in ln]
+        check(any("{host}" in ln for ln in cookie_dom_lines),
+              "catalog/render.py's sidecar --cookie-domain is not derived from the route host. A "
+              "zone-wide literal (`.{ZONE}`) broadcasts every app's session cookie across the "
+              "whole zone — see assertion (4)")
+        check(not any(re.search(r"--cookie-name=_oauth2_proxy_[a-z0-9]", ln) for ln in cookie_name_lines),
+              "catalog/render.py hardcodes a literal service name into --cookie-name. Whatever "
+              "service that is, every OTHER service now renders the same constant and they all "
+              "collide. Derive the name instead")
+
+# ============================================================ (6) PROVE THE SHIPPED RENDERER
+# Once the tree is compliant, (1)-(4) have nothing left to catch and their passing is evidence
+# of nothing. This runs the SHIPPED renderer as a SUBPROCESS on two synthetic ServiceDefinitions
+# and asserts the property directly, so the guard stays falsifiable.
+def _sd(name: str, port: int) -> str:
+    return yaml.safe_dump({
+        "name": name,
+        "kind": "container",
+        "source": {"image": "example/app:1"},
+        "exposure": "internal",
+        "ingressRoutes": [{"host": f"{name}.home.keithjackson.dev", "port": port}],
+        "auth": {"mode": "forward-auth", "groups": ["platform-admins"]},
+    }, sort_keys=False)
+
+
+def _sidecar_args(rendered: str) -> list:
+    for doc in yaml.safe_load_all(rendered):
+        if not isinstance(doc, dict) or doc.get("kind") != "Deployment":
+            continue
+        for c in iter_containers(doc):
+            if c.get("name") == "oauth2-proxy":
+                return list(c.get("args") or [])
+    return []
+
+
+if render_py.is_file():
+    fixtures = {"cookiefixa": 8111, "cookiefixb": 8112}
+    rendered_args: dict[str, list] = {}
+    with tempfile.TemporaryDirectory() as td:
+        for svc, port in fixtures.items():
+            f = Path(td) / f"{svc}.yaml"
+            f.write_text(_sd(svc, port))
+            proc = subprocess.run([sys.executable, str(render_py), str(f)],
+                                  capture_output=True, text=True)
+            check(proc.returncode == 0,
+                  f"the shipped catalog/render.py failed on a synthetic forward-auth fixture "
+                  f"({svc}): rc={proc.returncode} {proc.stderr.strip()[:200]}")
+            if proc.returncode == 0:
+                rendered_args[svc] = _sidecar_args(proc.stdout)
+
+    names, domains = {}, {}
+    for svc, args in rendered_args.items():
+        names[svc] = flag_from_args(args, "cookie-name")
+        domains[svc] = flag_from_args(args, "cookie-domain")
+        check(bool(args),
+              f"the shipped renderer produced no oauth2-proxy sidecar for synthetic forward-auth "
+              f"fixture {svc} — the sidecar is how a forward-auth app is gated at all (R-IDP-02)")
+        check(names[svc] not in (None, DEFAULT_COOKIE_NAME),
+              f"the shipped renderer gave synthetic fixture {svc} the cookie name "
+              f"{names[svc]!r}. A newly-onboarded forward-auth app would therefore default "
+              f"straight into the fleet collision — which is exactly how five services acquired "
+              f"this defect without anyone making a decision")
+        check(domains[svc] == f"{svc}.home.keithjackson.dev",
+              f"the shipped renderer scoped synthetic fixture {svc}'s cookie to "
+              f"{domains[svc]!r} rather than to its own host — a newly-onboarded app would "
+              f"broadcast its session cookie across the whole zone")
+        check(any(isinstance(a, str) and a.startswith("--whitelist-domain=.") for a in args),
+              f"synthetic fixture {svc} lost its ZONE-WIDE --whitelist-domain. That flag governs "
+              f"permitted post-login REDIRECT targets, not cookie scope; narrowing it with the "
+              f"cookie domain breaks the sign-in bounce")
+
+    if len(rendered_args) == 2:
+        a, b = sorted(rendered_args)
+        check(names[a] != names[b],
+              f"the shipped renderer gave two DIFFERENT synthetic services the SAME cookie name "
+              f"({names[a]!r}). Whatever the committed goldens currently look like, the generator "
+              f"will collide every forward-auth app it renders from now on")
+        check(domains[a] != domains[b],
+              f"the shipped renderer gave two different synthetic services the same cookie domain "
+              f"({domains[a]!r}) — the cookies would share a scope even with distinct names")
+
+# ============================================================ VERDICT
+inv = " | ".join(f"{i['key']} -> {i['effective_name']} @ {i['cookie_domain'] or '<host>'}"
+                 for i in sorted(instances, key=lambda x: x["key"]))
+print(f"oauth2-proxy instances audited ({len(instances)}): {inv}")
+if FAILURES:
+    print(f"\nCHECK FAIL (oauth2-proxy-cookie-names-distinct): {len(FAILURES)} failed, "
+          f"{CHECKS - len(FAILURES)} passed", file=sys.stderr)
+    for f in FAILURES:
+        print(f"  FAIL {f}", file=sys.stderr)
+    raise SystemExit(1)
+print(f"check OK: oauth2-proxy-cookie-names-distinct — {CHECKS} assertions (R-IDP-02, R-IDP-06)")
+PY
@@ -129,6 +129,22 @@ KNOWN_EXTERNAL_MONITORS: dict[str, str] = {
        "a PodMonitor, so it resolves against POD ports, not Service ports — the Flux controller "
        "Deployments (and their `http-prom` container port) are installed by the Flux bootstrap "
        "manifests, not authored here. Live on prod: job flux-system/flux-system, 6 targets up.",
+    "platform/oauth2-proxy/podmonitor.yaml:oauth2-proxy-sidecars":
+        "a PodMonitor, so it resolves against POD ports, not Service ports. Unlike the other "
+        "entries here the target is NOT chart-owned — the `metrics` (44180) container port is "
+        "authored in this repo, on the oauth2-proxy sidecar in apps/wave-0/{jackett,qbittorrent,"
+        "pihole}.yaml and apps/wave-1/{radarr,sonarr}.yaml. It is unauditable here because NO "
+        "Service exposes 44180 BY DESIGN: adding one would mean changing catalog/render.py and "
+        "every apps/wave-* render and golden, to publish an endpoint only Prometheus should dial. "
+        "A PodMonitor needs no Service, which is the whole reason it was chosen. What assertion "
+        "(2) would have checked was instead MEASURED on prod 2026-10-01, read-only: the pods "
+        "declare ports [('http', 4180), ('metrics', 44180)], the selector label "
+        "homelab.keithjackson.dev/auth-mode=forward-auth is on exactly those five gate pods, and "
+        "port-forwarding 44180 returns http=200 with 10208 bytes of oauth2_proxy_* metrics. "
+        "CAVEAT, deliberately recorded: those targets are DOWN on prod, because the app "
+        "namespaces' default-deny NetworkPolicy drops the scrape from ns monitoring (same bug and "
+        "fingerprint as platform/longhorn/networkpolicy-metrics.yaml documents). The port name "
+        "resolves; the network path does not yet.",
 }

 MONITOR_KINDS = ("ServiceMonitor", "PodMonitor")
@@ -76,11 +76,14 @@ spec:
        - --allowed-group=platform-admins
        - --skip-provider-button=true
        - --redirect-url=https://paas-demo.home.keithjackson.dev/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_paas-demo
+        - --cookie-domain=paas-demo.home.keithjackson.dev
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -101,6 +104,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -99,11 +99,14 @@ spec:
        - --allowed-group=media-admins
        - --skip-provider-button=true
        - --redirect-url=https://sonarr.home.keithjackson.dev/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_sonarr
+        - --cookie-domain=sonarr.home.keithjackson.dev
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -124,6 +127,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -99,11 +99,14 @@ spec:
        - --allowed-group=media-admins
        - --skip-provider-button=true
        - --redirect-url=https://sonarr.home.keithjackson.dev/oauth2/callback
-        - --cookie-domain=.${BASE_DOMAIN}
+        - --cookie-name=_oauth2_proxy_sonarr
+        - --cookie-domain=sonarr.home.keithjackson.dev
        - --whitelist-domain=.${BASE_DOMAIN}
        - --cookie-secure=true
+        - --cookie-refresh=2m30s
        - --pass-access-token=true
        - --set-xauthrequest=true
+        - --metrics-address=0.0.0.0:44180
        env:
        - name: OAUTH2_PROXY_CLIENT_ID
          valueFrom:
@@ -124,6 +127,9 @@ spec:
        - name: http
          containerPort: 4180
          protocol: TCP
+        - name: metrics
+          containerPort: 44180
+          protocol: TCP
        resources:
          requests:
            cpu: 25m
@@ -0,0 +1,283 @@
+---
+date: 2026-09-30
+branch: feat/node-health-render-device
+title: The Intel render device is now self-re-measuring, not a fact someone checked once
+---
+
+## What changed
+
+### Why this probe has value at all
+
+**Nothing else in the platform would notice the device going missing.** Plex exposes no
+ServiceMonitor, so no alert fires on a silent fall back to software transcoding. And the
+manifest side mounts the device with `hostPath type: DirectoryOrCreate`, which means a node
+*without* the device mounts an **empty directory** — Plex starts, runs, and degrades quietly
+rather than failing loudly. There is no crash loop, no failed mount, no alert: just a hot CPU.
+This check is the only thing in the system that would surface it, which is the actual
+justification for the role carrying it.
+
+### The change
+
+`roles/node_health` grew a graphics / render-device probe. Nothing was installed, no module
+was loaded, no group was changed, nothing was rebooted: the prod nodes already have a working
+Intel render device (measured by hand on all three on 2026-09-30) and the point of this change
+is that the measurement now **re-runs forever** instead of decaying silently in a handoff note.
+
+The failure this exists to catch is silent by construction. Plex hardware-transcodes through
+`/dev/dri/renderD128`; if that node stops existing — node rebuild, a BIOS change that disables
+the onboard IGD, a kernel that binds a different DRM driver, a driver regression — Plex falls
+back to software transcoding on its own and the only symptom is a hot CPU.
+
+Four files:
+
+- `files/node_health_collect.py` — new `collect_graphics()`, wired into the **unprivileged**
+  pass (`main()`). Everything it reads is world-readable: `lstat()` on `/dev/dri/*`,
+  `grp.getgrgid()` for the owning group name, the `/sys/class/drm/card*/device/driver` symlink
+  target for the bound driver, `/sys/bus/pci/devices/*/class` for display-class (PCI base class
+  `0x03`) devices whether or not a driver is bound, `/sys/class/drm/card*/engine/*` for the
+  `vcs*`/`vecs*` media engines, `/proc/modules`, `/proc/cmdline`, and
+  `/sys/module/{i915,xe}/parameters/enable_guc`. `lspci` and `vainfo` were added to the
+  existing tool probe; both are optional and neither is depended on.
+- `tasks/assess.yml` — one shared derive task plus three checks: `graphics.render_device`
+  (PASS/FAIL/SKIP), `graphics.drm_driver` (PASS/WARN/FAIL/SKIP), `graphics.media_stack`
+  (INFO/SKIP).
+- `defaults/main.yml` — `node_health_graphics_expected_driver: i915`, so the one comparison the
+  check makes is data rather than task logic.
+- `templates/summary.txt.j2` — a `GRAPHICS / RENDER DEVICE` section, and
+  `roles/node_health/README.md` gained a section explaining the design and two new "known
+  limits".
+
+Three design rulings worth stating, because each is the kind of thing a later edit would get
+wrong:
+
+1. **The render node's GID is reported, never asserted.** It read `root:render`, gid **992**,
+   mode `0660` on all three prod nodes. The check does *not* compare against `992`. `render` is
+   a **systemd dynamic** allocation (`systemd-sysusers`), not a Debian static gid from
+   `base-passwd`, so a rebuild can legitimately move it — and Plex's container entrypoint stats
+   the device and joins whichever group owns it, so drift genuinely is not a fault. The
+   observed group name and number appear in the detail line, which makes drift *visible*
+   without making it an error. A hard `!= 992` would be a false alarm waiting for a reinstall.
+2. **The PCI scan is deliberately separate from the DRM scan.** `/sys/class/drm/card*` exists
+   only once a driver has *bound*, so a GPU with nothing attached has no `card0` and would
+   otherwise be indistinguishable from a machine with no GPU. Scanning the PCI bus separately
+   is what lets "no display device at all" be an honest **SKIP** (a VM guest — the same
+   convention `thermal.max` already uses for "no sensor on this hardware") while "a GPU is on
+   the bus and there is still no render node" is a **FAIL**.
+3. **GuC/HuC and host VAAPI userspace are INFO and must stay INFO.** GuC/HuC off
+   (`i915.enable_guc=0`) is an accepted state here and costs low-power H.264 encode *quality*
+   only. No host `vainfo` is not a defect at all — Plex ships its own libva and media driver
+   inside its container. Grading either would put a permanent WARN on a chosen configuration,
+   or invite someone to `apt install vainfo` on a node to clear a health check, and this role
+   installs nothing.
+
+`xe` bound instead of `i915` is a **WARN**, not a FAIL: the device still works, but the two
+drivers do not share a userspace media stack. On these nodes `xe.ko.xz` exists on disk and is
+bound to nothing.
+
+## What I verified
+
+### The real prod run (read-only, all three nodes)
+
+`--list-hosts` first, every time. It named exactly three hosts and nothing else:
+
+    play #1 (k3s_prod): Inspect node hardware health (read-only)
+        pattern: ['k3s_prod']
+        hosts (3): k3s-prod-3  k3s-prod-2  k3s-prod-1
+
+Then `--syntax-check` (rc 0), then the run:
+
+    ansible-playbook playbooks/node-health.yml -e node_health_hosts=k3s_prod --limit k3s_prod
+
+    k3s-prod-1 : ok=51  changed=3  unreachable=0  failed=0  skipped=1
+    k3s-prod-2 : ok=51  changed=2  unreachable=0  failed=0  skipped=1
+    k3s-prod-3 : ok=51  changed=2  unreachable=0  failed=0  skipped=1
+    PLAY rc=0
+
+(The `changed=` counts are the *controller-side* artifact writes under
+`infra/ansible/_artifacts/node-health/`, delegated to localhost. Nothing on any node changed —
+the collector task is `changed_when: false` and the role touches nothing.)
+
+All three nodes graded **PASS / PASS / INFO** on the new checks, 0 FAIL overall:
+
+    k3s-prod-1  VERDICT  0 FAIL · 1 WARN · 18 PASS · 2 SKIP · 6 INFO
+    k3s-prod-2  VERDICT  0 FAIL · 3 WARN · 15 PASS · 3 SKIP · 6 INFO
+    k3s-prod-3  VERDICT  0 FAIL · 2 WARN · 17 PASS · 2 SKIP · 6 INFO
+
+    [PASS] graphics.render_device — DRM render node present for hardware transcoding (/dev/dri/renderD*)
+    [PASS] graphics.drm_driver   — DRM driver bound to the display device (expected i915)
+    [INFO] graphics.media_stack  — Media engines, GuC/HuC firmware and host VAAPI userspace (informational)
+
+(The pre-existing WARNs and SKIPs are other checks — SMART, memattr, resolvers — and are not
+mine; none of them moved.)
+
+The new report section, **byte-identical on all three nodes**:
+
+    --------------------------------------------------------------------------------
+     GRAPHICS / RENDER DEVICE (the Plex hardware-transcoding path)
+    --------------------------------------------------------------------------------
+     pci  0000:00:02.0 00:02.0 VGA compatible controller [0300]: Intel Corporation CoffeeLake-S GT2 [UHD Graphics 630] [8086:3e98] (rev 02)
+                  driver i915
+     drm  card0        driver i915     at 0000:00:02.0  engines bcs0 rcs0 vcs0 vecs0
+     dev  card0        character special  mode 0660  root:video gid 44
+     dev  renderD128   character special  mode 0660  root:render gid 992
+     GuC/HuC     NOT READ — /sys/module/i915/parameters/enable_guc is not
+                  readable without root (i915 module parameters are mode 0400)
+                 no i915/xe parameter on the kernel cmdline
+     modules     amdgpu=false, i915=true, nouveau=false, nvidia=false, radeon=false, xe=false (/proc/modules — corroboration only; the bound-driver symlink above decides)
+     vainfo      NOT installed on this host — not a defect: Plex ships its own libva/media driver inside its container
+
+That reproduces every item of the 2026-09-30 hand measurement independently: `renderD128`
+character special, `root:render`, **gid 992**, mode `0660`; `card0` `root:video` gid 44 mode
+`0660`; `i915` bound; `xe` **not** loaded and bound to nothing; `[8086:3e98] (rev 02)`
+CoffeeLake-S GT2 UHD Graphics 630 at `00:02.0`; `vcs0` and `vecs0` live; no host `vainfo`.
+
+**One measured correction to my own phase-1 assumption:** `lspci` **is** present on all three
+prod nodes (`/usr/bin/lspci` — Debian ships `pciutils` by default). I had written that it was
+absent. It is still *treated* as optional, because `pciutils` is genuinely not declared in
+`node_baseline_packages` and depending on an undeclared package would be wrong — but the
+comments and README that asserted it was missing were factually wrong and are now corrected.
+Its presence is why the `pci` line above carries a readable device name rather than bare IDs.
+
+### The gate
+
+    DEV_KUBECONFIG=/dev/null ENGINE=<shim> ./infra/devcontainer/dev.sh run bash -c './ci/gate.sh'
+    GATE rc=3
+
+**rc 3 — UNDETERMINED. NOT green.** Nothing failed, and nothing is certified either. Exactly
+three stages did not run, all three the known local ceiling from a worktree, and **no fourth
+skip**:
+
+    [skip] spec-immutability  spec repo not present at /work/../home-network-k3s (set SPEC_REPO)
+    [skip] gitea-settings     SETTINGS_API_TOKEN is not set — cannot read the live Gitea rule
+    [skip] conformance        no cluster configured (CLUSTER_AVAILABLE!=true) — L2/L3 NOT RUN
+
+    [ ok ] decisions-log: pass      [ ok ] lint: pass       [ ok ] l0: pass
+    [ ok ] render-golden: pass      [ ok ] l1: pass         [ ok ] approval-policy: pass
+
+Both CLAUDE.md §3 worktree workarounds were needed and both worked: `DEV_KUBECONFIG=/dev/null`
+(the credential in the main checkout was **not** copied) and an `ENGINE` shim that bind-mounts
+the main checkout's `.git` read-only. No `tracked_files()` traceback appeared, which is how I
+know the shim took effect rather than the git stages silently degrading.
+
+`ansible-lint` was **not** run, because **`ci/` does not invoke it anywhere** — I grepped
+`ci/` and `.gitea/` and the lint stage is `yamllint -c ci/yamllint.yaml` only. I am not
+claiming to have matched a lint convention that does not exist.
+
+### The CI gate on the PR — a green tick that is exit 3, read from the log
+
+The required context `gate / gate (pull_request)` reported **`success`, "Successful in 1m40s"**
+on commit `a78f010`. That description matters: a *skipped* Gitea job reports `success` with
+"Has been skipped", so this one genuinely ran, and 1m40s matches CLAUDE.md §2's ~1m35s.
+
+**But the green tick does NOT mean `gate.sh` returned 0.** `.gitea/workflows/gate.yml:138` is
+`run: ./ci/gate.sh || [ "$?" -eq 3 ]` — the job tolerates rc 3 by design. So I fetched the job
+log (`/actions/runs/173/jobs/0/logs`; the `/api/v1/.../tasks/{id}/logs` endpoints 404 on this
+Gitea, the web route serves it) and read the gate's own headline:
+
+    [fail] MERGE GATE: UNDETERMINED — greenness was NOT ESTABLISHED.
+    [fail]   Stages that did not run: spec-immutability
+
+**So CI is also rc 3, with exactly ONE stage skipped — `spec-immutability`.** The green Gitea
+check is the tolerance swallowing that 3, not a rc 0.
+
+**CI verified two things my local run could not, and both came back green:**
+
+    [ ok ] stage gitea-settings: pass
+    [ ok ]   enable_push = false          enable_status_check = true
+    [ ok ]   required_approvals = 1       block_admin_merge_override = true
+    [ ok ]   block_on_rejected_reviews = true   dismiss_stale_approvals = true
+    [ ok ]   runners online: 1 (>= 1)
+    [ ok ] Gitea settings match the declaration.
+    [ ok ] stage approval-policy: pass
+    [ ok ] no architecturally/security-relevant paths touched — the GATE is the approver.
+
+My local run **skipped** `gitea-settings` (no `SETTINGS_API_TOKEN`) and could not give
+`approval-policy` a PR number. CI has the secret and `REQUIRE_GITEA_SETTINGS: "true"`, so a skip
+there would have failed the job. That is measured confirmation that the live Gitea matches
+`infra/gitea/expected-branch-protection.json`, and that `infra/ansible/**` needs no human
+reviewer. `lint` passed over 405 files. `conformance` SKIPPED in CI too
+(`CLUSTER_AVAILABLE != true`), so L2/L3 remains unverified there as well.
+
+Corroborated independently: `GET /repos/.../branches/main` →
+`protected=true user_can_push=false enable_status_check=true required_approvals=1`,
+contexts `["gate / gate (pull_request)"]`; `GET /admin/actions/runners` → `incluster-dind
+status=online`. And note the commit's **combined** state is `pending`, not success, because
+`gate / conformance (pull_request)` sits at "Blocked by required conditions" — that context is
+not in the required list, so it does not block, but the PR is not "all green" either.
+
+The gate's `lint` stage passed with one `line too long` **warning** on
+`roles/node_health/tasks/assess.yml:189` — that is pre-existing, in the SMART block, well
+above my insertion point, and warnings do not fail the stage.
+
+### Offline work (before the lock was released)
+
+- `python3 -m py_compile` on the collector: clean.
+- `collect_graphics()` executed for real against this workstation's hardware (an unrelated
+  machine, Raptor Lake-S UHD Graphics + an NVIDIA discrete GPU) and returned the whole
+  structure with no exception, correctly reporting two cards, two render nodes, the bound
+  driver per card, the `vcs0/vcs1/vecs0` media engines, and `enable_guc` as *not readable
+  without root* rather than as a fabricated `0`.
+- That last point is a real finding and changed the design: `/sys/module/i915/parameters/*` is
+  **mode 0400**, so an unprivileged pass cannot read `enable_guc`. It now reports
+  `NOT READ — <reason>` and `/proc/cmdline` is read as the world-readable second source. An
+  earlier draft would have silently reported a null.
+- Two other bugs found and fixed by running it rather than reasoning about it:
+  `/dev/dri/by-path` is a **directory** and was being `lstat()`ed as a device; and
+  `/sys/class/drm` carries one entry per *connector* (`card0-HDMI-A-1`), which had to be
+  excluded from the card scan.
+- All three checks and the new template section were rendered through a standalone Jinja
+  harness (Ansible is not on the host — CLAUDE.md §5) against **seven** fact sets: the real
+  workstation; GPU present with the render node removed; no display device at all; `xe` bound
+  in place of `i915`; a display device with **no** driver bound; `renderD128` present but not a
+  character device; and a shape with neither `lspci` nor `vainfo`. Zero template errors, and
+  the statuses came out as designed — `FAIL/PASS/INFO`, `SKIP/SKIP/SKIP`, `PASS/WARN/INFO`,
+  `FAIL/FAIL/INFO`, `FAIL/PASS/INFO` respectively. The harness supplies stand-ins for the
+  Ansible-only filters (`bool`, `dict2items`, `flatten`, `zip`, `regex_replace`) and matches
+  Ansible's `trim_blocks=True`.
+  **Read that for exactly what it is: evidence that the LOGIC is right, and no evidence
+  whatsoever that the HARDWARE DETECTION is.** Six of those seven fact sets were hand-mutated
+  dictionaries, not machines. They are not hardware tests and must not be written up as if
+  they were.
+
+## What I did NOT verify
+
+- **The FAIL path has never fired on a real node.** Every prod node has a working render
+  device, and this role is read-only, so a device cannot be taken away to prove the check
+  catches its absence. The failure branches were exercised against **synthetic fact sets
+  only** (see above). I am not claiming the failure path is tested against hardware.
+- **`xe` detection is likewise untested against reality.** No node here has `xe` bound — the
+  module exists on disk as `xe.ko.xz` and is bound to nothing — so the WARN branch was only
+  ever driven by a mutated fact set. If `xe` ever does bind, the *shape* of
+  `/sys/class/drm/card*` under it is something I have assumed matches i915's (same
+  `device/driver` symlink, same `engine/*` directories). That assumption is unverified.
+- **The media-engine sysfs path is confirmed on these nodes but not proven portable.**
+  `vcs0`/`vecs0` were read from `/sys/class/drm/card0/engine/*` on all three prod nodes, so it
+  is measured here. I did **not** confirm against kernel source that no other kernel version or
+  driver publishes engines elsewhere. If one does, the check reports "none exposed" — an INFO
+  line, not a false verdict, which is why this failure mode is tolerable.
+- **`enable_guc` is still unread on prod.** The report says `NOT READ — mode 0400`, which is
+  honest but means I have **not** confirmed by measurement that GuC/HuC is off on these nodes;
+  I only know the kernel cmdline carries no `i915.enable_guc=` override. The brief's ground
+  truth says `i915.enable_guc=0`; my check does not corroborate that value, it declines to
+  guess it. Reading it would need root, and it is INFO either way, so I left it alone.
+- **Not a transcode test.** These checks prove the render *device* exists, is a character
+  device, and which driver owns it. They do **not** prove a transcode succeeds — that needs a
+  VAAPI round trip inside Plex's own container with Plex's own libva, and is out of scope for a
+  node-level check that installs nothing.
+- **Greenness was never established, locally or in CI.** Both runs returned **rc 3**. My local
+  run skipped three stages (`spec-immutability`, `gitea-settings`, `conformance`); CI skipped
+  one (`spec-immutability`) and its own headline says `MERGE GATE: UNDETERMINED`. **I am not
+  describing this PR as green**, and the green Gitea tick must not be read as one: the workflow
+  line is `./ci/gate.sh || [ "$?" -eq 3 ]`, so it is a tolerated 3.
+- **L2/L3 conformance did not run anywhere** — `CLUSTER_AVAILABLE != true` in both. Whatever
+  runtime requirements this touches remain UNVERIFIED; only the L0/L1 offline subset passed.
+- **`spec-immutability` ran in neither.** The sibling spec repo is checked out in neither my
+  worktree nor the CI workspace, so nothing confirms this change does not drift from the spec.
+  It is the expected, tolerated skip — but it is still a skip, not a pass.
+- **No claim about the live platform beyond what was printed.** Per CLAUDE.md §2b this repo is
+  not evidence about prod; the only prod facts I assert are the ones quoted verbatim above.
+  Branch protection and runner state I did query directly, and the CI `gitea-settings` stage
+  passed — so those two I assert on measurement, not on this file's prose.
+- **The stale dev inventory was left alone deliberately.** `hosts.ini` still defines the
+  `k3s-dev-*` groups as the default target of most playbooks and the dev cluster no longer
+  exists. That is real, it is out of this change's scope, and it is unclaimed work.
@@ -0,0 +1,289 @@
+---
+date: 2026-09-30
+branch: feat/oauth2-proxy-auth-failure-visible
+title: The Jackett auth failure was invisible because the gates emit nothing — alert on the gap, not on a metric that does not exist
+---
+
+## What changed
+
+Jackett's UI broke in the way that costs the most: **HTTP 200 and a plausible page**. The
+operator's report named the property that matters — *"reloading the page never fixes the issue and
+the user is never re-prompted for login"*. Two sibling branches fix the auth defects. This one is
+about the fact that **no signal existed anywhere**, and it turns out that is not an oversight in
+the alert rules. It is an absence of data.
+
+**The measurement that decided the whole design** (prod, live Prometheus, 2026-09-30):
+
+    count by (__name__) ({__name__=~"oauth2_proxy.*"})
+      oauth2_proxy_requests_in_flight => 2
+
+Two series. Both from the **shared** forward-auth Deployment in the `oauth2-proxy` namespace.
+That is every oauth2-proxy metric this cluster holds. The six per-app gate sidecars — crafty,
+jackett, pihole, qbittorrent, radarr, sonarr — publish **nothing at all**: they run with
+`--http-address=0.0.0.0:4180` and no `--metrics-address`, so oauth2-proxy's metrics listener never
+starts; they declare only a `http:4180` container port; and no ServiceMonitor or PodMonitor in the
+cluster selects them.
+
+So the alert the brief asked for — *"a gated app is serving a sustained rate of auth redirects to
+a user who is authenticated"* — **is not writable**, and writing it anyway would have produced
+exactly the artefact this repo keeps getting burned by: a rule that is present, evaluating,
+permanently silent, and indistinguishable from health. It was not written.
+
+What was written instead alerts on the condition that made the incident invisible, from families
+that **are** scraped:
+
+- **`platform/oauth2-proxy/auth-visibility-alerts.yaml`** — `OAuth2ProxyGateUnobserved`:
+
+      count by (namespace) (kube_pod_container_info{container="oauth2-proxy"})
+        unless
+      count by (namespace) (up{container="oauth2-proxy"})
+
+  `for: 1h`, `severity: warning`. `kube_pod_container_info` is kube-state-metrics (a declared
+  owner in `alert-metrics-have-scrape-targets`) and sees the container whether or not it exports
+  anything; `up` is UNIVERSAL and carries a `container` label for any scrape target, so the rule
+  needs no advance knowledge of what job name a future per-app scrape will use. Nothing in it
+  touches `oauth2_proxy_*`, so it cannot be inert for the reason it is reporting. **It returns six
+  namespaces against live prod right now** and excludes `oauth2-proxy`, which is scraped. It
+  clears per-namespace as each gate gains a real scrape.
+- **`platform/oauth2-proxy/auth-visibility-alerts.promtool-test.yaml`** — four states pinned:
+  fires for an unscraped namespace, never for the scraped one, honours `for: 1h` (pending at 59m),
+  and **stops** once a scrape appears under a *different* job name. Series shapes copied off live
+  prod, not invented.
+- **`runbooks/oauth2-proxy-auth-failures.md`** — how to tell "the gate refused me" from "the group
+  check refused me" from "the app refused me" from "the XHR silently got HTML", with the
+  2026-09-30 jackett trace as the worked example. Registered in `runbooks/kustomization.yaml`'s
+  `configMapGenerator` and listed in `runbooks/README.md`.
+- **`runbooks/README.md`** — a new row in *"Alerts you should not trust the silence of"* for the
+  per-app sidecars, plus the measured reason.
+- **`platform/oauth2-proxy/README.md`** — what this component can and cannot see, and the exact
+  three-part change that would close the gap.
+
+**Why the obvious alternative was rejected.** An alert on the *shared* proxy's
+`oauth2_proxy_requests_total{code="302"}` ratio is the one a reader expects to see here. It would
+need a new `SCRAPE_OWNERS` entry in `conformance/**` (a sensitive path needing human approval),
+and — decisively — **it would not have fired on this incident**: the shared instance is not in the
+data path for any sidecar-gated app, and its current pods export no `oauth2_proxy_requests_total`
+at all, having proxied nothing since they started (the counter exists in the TSDB only from an
+older ReplicaSet, `oauth2-proxy-78449b5657-*`, with `code` labels `302` and `401` and no `202`).
+Shipping it would have looked like coverage of the reported failure and been none.
+
+## What I verified
+
+Everything below was measured against live prod through the pinned container, read-only. No
+`apply`, no `edit`, no `patch`, no `flux reconcile`.
+
+- **The full oauth2-proxy metric inventory.** `count by (__name__) ({__name__=~"oauth2_proxy.*"})`
+  → `oauth2_proxy_requests_in_flight => 2`. `count by (__name__) ({job="oauth2-proxy"})` returns
+  go_*/process_*/promhttp_* plus that one family and nothing else. **Only the shared instance is
+  scraped**: `up{job="oauth2-proxy"}` → 2 targets, `10.0.1.189:44180` and `10.0.2.47:44180`, both
+  `1`, `namespace="oauth2-proxy"`.
+- **The per-app sidecars are unscraped, and unscrapable as configured.** Every pod carrying an
+  `oauth2-proxy` container, with its declared ports: the two shared replicas have
+  `[('http',4180),('metrics',44180)]`; crafty, jackett, pihole, qbittorrent, radarr and sonarr all
+  have `[('http',4180)]` only. The live jackett sidecar's args contain no `--metrics-address`.
+  `kubectl get servicemonitors,podmonitors -A` lists 16 ServiceMonitors and 5 PodMonitors; the
+  only oauth2-proxy entry is `oauth2-proxy/oauth2-proxy`.
+- **No existing alert is inert *for this reason*, because no alert references these metrics at
+  all.** `grep -rn oauth2_proxy --include=*.yaml` over the tree hits only cookie-name strings in
+  `platform/crafty/`. The 39 live PrometheusRules include `platform-health-identity`, whose four
+  rules are about Keycloak readiness and the client reconciler and are all keyed on `kube_*`
+  (scraped, fine). **The inertness here is the absence of a rule, not a broken one** — which is
+  why the runbook README's inert-alerts table gains a row phrased like the etcd one.
+- **Keycloak cannot express it either.** `count by (__name__) ({__name__=~"keycloak_.*"})` →
+  empty, and empty over `[30d]`. The `keycloak-service` scrape yields JVM/Agroal/JGroups metrics
+  only; there are no user-event or login counters, so "a burst of re-logins for one client" is not
+  observable from the IdP side.
+- **`oauth2_proxy_requests_total` has existed historically but not on the current pods.**
+  `count_over_time(oauth2_proxy_requests_total[30d])` returns three series, all from
+  `oauth2-proxy-78449b5657-*` (a prior ReplicaSet): `code="302"` on both replicas and `code="401"`
+  on one, ~634 samples each (~5h of a 30d window). **No `code="202"` ever.**
+- **The alert expression fires against live data.** Run verbatim on prod:
+  `count by (namespace) (kube_pod_container_info{container="oauth2-proxy"}) unless count by
+  (namespace) (up{container="oauth2-proxy"})` → `pihole, jackett, crafty, qbittorrent, sonarr,
+  radarr`, value 1 each; `oauth2-proxy` correctly absent.
+- **The log signal exists and is collected.** Loki (`loki-0`, SingleBinary) and a 3-node Alloy
+  DaemonSet are running, and Alloy labels pod logs with `namespace`/`container`. LogQL over 24h,
+  `{container="oauth2-proxy"}`: crafty 12036 lines, jackett 11669, oauth2-proxy 34559, pihole
+  11520, qbittorrent 11520, radarr 11763, sonarr 11520. Filtered to `|= "No valid authentication
+  in request"`: **crafty 81, jackett 39, radarr 67**; to `|= "AuthSuccess"`: **crafty 1, jackett
+  4, radarr 1**. Those ratios (81:1, 10:1, 67:1) are the incident's signature and are in the
+  runbook as the thing to compare against.
+- **Why `kubectl logs` hid it.** The jackett sidecar's stdout is ~11.5k lines/day and almost all
+  of it is `GET "/ping" ... 200 2` from the kubelet probe, twice every 15s. A `--tail=40` shows
+  forty `/ping` lines and looks perfectly healthy. The runbook's commands all filter.
+- **Runbook registration, both halves.** `- oauth2-proxy-auth-failures=oauth2-proxy-auth-failures.md`
+  is in `runbooks/kustomization.yaml`'s `configMapGenerator.files`, and the page is listed under
+  *Identity* in `runbooks/README.md`. The page clears the L0 quality floor (the four operator
+  sections, >1500 bytes, mentions `dev.sh` and `--context k3s-prod`), and the alert's
+  `runbook_url` points at exactly that name.
+- **Gate:** run in the pinned toolbox from this worktree with the `ENGINE` shim. Exit code and
+  per-stage outcomes in the next section — read them there, not here.
+
+## What I did NOT verify
+
+- **The alert has never fired in Alertmanager.** Nothing is applied; the PrometheusRule exists
+  only in git on this branch. What is proven is that its expression returns the expected vector
+  against live prod data and that promtool agrees about its behaviour over synthetic series.
+  **Merging is what deploys it.**
+- **`OAuth2ProxyGateUnobserved` would NOT have fired *during* the 2026-09-30 jackett incident in
+  any useful sense** — it describes a standing condition that has been true since the sidecars
+  were deployed, so it would have been firing before, during and after. That is deliberate and it
+  is the honest claim: it is a **gap tracker**, not an incident detector. It would have meant that
+  on 2026-09-30 an operator asking "why did nothing alert?" had the answer already on the board
+  instead of discovering it during the investigation. **No alert on this platform could have
+  detected the incident itself, and none can today.**
+- **The alert that the brief actually asked for does not exist, and cannot be written yet.** The
+  sustained-redirect-ratio rule needs `oauth2_proxy_requests_total` from the *per-app* sidecars.
+  **Enabling that requires editing files owned by the two sibling branches** — the sidecar args in
+  `apps/wave-0/jackett.yaml` (Mira Vance) and `apps/wave-0/{qbittorrent,pihole}.yaml`,
+  `apps/wave-1/{radarr,sonarr}.yaml` (Theo Bright) — so it was **not attempted here**. The change
+  is three things per app, together: `--metrics-address=0.0.0.0:44180`, a `metrics` containerPort
+  plus Service port, and a ServiceMonitor. Once any one app has them, the rule belongs beside
+  `OAuth2ProxyGateUnobserved`, and that app's namespace drops out of it automatically.
+- **The Loki alerting path is inert and I did not fix it.** Loki *has* a ruler stanza and the
+  chart's `loki-sc-rules` sidecar is running with `LABEL=loki_rule`, `FOLDER=/rules` — so a
+  ConfigMap-shipped LogQL rule would be delivered. But `GET /loki/api/v1/rules` answers **"unable
+  to read rule dir /var/loki/rules/fake: no such file or directory"** (the ruler reads a different
+  directory from the one the sidecar writes), and the rendered config contains **no
+  `alertmanager_url`** — `grep -i 'alertmanager\|rule_path'` over the `loki` ConfigMap returns
+  nothing. So a LogQL alert on `|= "No valid authentication in request"` — which *would* have
+  caught this incident, and is the single highest-value follow-up here — **would evaluate and
+  notify nobody**. Wiring it means `rulerConfig` (alertmanager_url + a matching rules directory)
+  in `platform/monitoring/loki.yaml`, which is not a file this brief gave me. **I did not ship an
+  inert rule ConfigMap to look productive.**
+- **Nothing was reproduced in a browser.** The 302-to-200-with-HTML mechanism is read off the
+  sidecar access log and the response sizes (`/UI/Dashboard 200 44379`), not observed in devtools.
+- **Loki retention is 72h** (`platform/monitoring/loki.yaml`), so the 24h ratios above are a
+  snapshot; anything older than three days is already gone. This is stated in the runbook because
+  it bounds what an operator can still investigate.
+- **The gate's L2/L3 conformance did not run** (`CLUSTER_AVAILABLE` unset), so no runtime
+  requirement touched by this branch is certified. See the exit code below — **rc 3 is
+  UNDETERMINED, not green.**
+
+### Follow-ups, in order of value
+
+1. **Wire the Loki ruler to Alertmanager** (`platform/monitoring/loki.yaml`). This is the only
+   route to an alert that would have caught the actual 2026-09-30 failure, and the data is
+   already being collected. The rule is one expression:
+   `sum by (namespace) (rate({container="oauth2-proxy"} |= "No valid authentication in request" [15m]))
+   > 5 * sum by (namespace) (rate({container="oauth2-proxy"} |= "AuthSuccess" [15m]))`.
+2. **Give one gate sidecar a metrics endpoint** and watch its namespace drop out of
+   `OAuth2ProxyGateUnobserved`. That is the cheapest proof the alert is a tracker and not a nag,
+   and it unlocks the ratio rule in Prometheus.
+3. **Consider whether `--set-xauthrequest` apps should get `--api-routes`** so API paths answer
+   401 instead of 302. That is the fix that would make this failure *loud* rather than merely
+   observable, and it belongs with whoever owns the sidecar args.
+
+---
+
+# Round 2 — 2026-10-01: the scrape got wired, and found a bug in this branch's own alert
+
+Added to the same PR (#29) rather than a fourth concurrent branch. PR #28 merged in the meantime
+and `main` moved three PRs ahead; `origin/main` is merged in here, and the `runbooks/kustomization.yaml`
+conflict was resolved by **keeping both** runbook entries (mine and `oauth2-proxy-session`).
+
+## What changed
+
+**`platform/oauth2-proxy/podmonitor.yaml` (new).** PR #28 gave the five fleet sidecars
+`--metrics-address=0.0.0.0:44180` and a named `metrics` port. No Service exposes 44180 on any of
+them, so a ServiceMonitor would have nothing to select — a **PodMonitor** targets pod ports directly
+and needs no renderer change. Selects on the pod label `homelab.keithjackson.dev/auth-mode:
+forward-auth`, which is on exactly the five fleet gate pods and absent from crafty's: the property,
+not an app list, so a sixth forward-auth app needs no edit.
+
+**`auth-visibility-alerts.yaml`: `up{container="oauth2-proxy"}` → `up{container="oauth2-proxy"} == 1`.
+This is a real bug fix, and it is the most important thing in this round.** A target that is
+discovered but failing still publishes `up`, with value **0**. The series exists, so
+`count by (namespace)` returned 1, `unless` removed the namespace, and `OAuth2ProxyGateUnobserved`
+would have **gone quiet while nothing was being scraped at all** — this branch's own disease,
+reproduced inside the alert written to cure it. Not hypothetical: the PodMonitor lands five targets
+whose scrape is dropped, so the bare form would have silently cleared on this very commit.
+
+**Two new unit-test scenarios** pinning that: a discovered-but-failing target must still fire, and a
+target that works and then breaks must re-fire rather than latch cleared.
+
+**Docs corrected rather than appended to.** The 2026-09-30 claim "the sidecars export no metrics at
+all" is now only half true and was rewritten in `runbooks/README.md`, `platform/oauth2-proxy/README.md`
+and the alert header, each showing the two-step history, because each step looked like the end of it.
+The runbook gains the policy-vs-no-target triage (step 1b) and the **radarr `/signalr` worked case**.
+
+**Deliberately NOT shipped: the ratio alert.** `oauth2_proxy_requests_total` now exists on the
+endpoint, but not in the TSDB, because the scrape is blocked. Writing the ratio rule today would make
+it inert for precisely the reason it was refused on 2026-09-30. It becomes writable the day a scrape
+succeeds, not the day a listener exists.
+
+## What I verified
+
+- **The endpoint serves — by payload, not by flag.** Port-forwarded to the jackett pod's 44180:
+  `http=200 bytes=10208`, exporting `oauth2_proxy_requests_in_flight`, `oauth2_proxy_requests_total`
+  (`code="200" 3`, `code="302" 2`) and `oauth2_proxy_response_duration_seconds`. `--metrics-address`
+  being in `args` and `/metrics` returning a body are different claims; both now hold.
+- **The scrape path is DROPPED, measured from the right place.** From grafana in ns `monitoring`
+  (beside Prometheus, same policy position) to the jackett pod IP: `:44180` → `wget: download timed
+  out`; `:4180` → **also** timed out; the shared proxy's `:44180` → 200 with `oauth2_proxy_*` present.
+  Cause, read from the objects: ns jackett's `default-deny` is `podSelector: {}`,
+  `policyTypes: [Ingress, Egress]`, with an `egress` rule and **no `ingress` key** (deny-all ingress);
+  the only allowance is CNP `jackett-allow-gateway-ingress`, `fromEntities: [ingress]` to port 4180.
+  The shared proxy is scrapable only because ns `oauth2-proxy` has **no NetworkPolicy at all**.
+- **No Service exposes 44180** on any of the five (enumerated).
+- **The pod label selector is exact:** `auth-mode=forward-auth` on all five fleet gate pods, absent
+  from crafty's. Crafty's sidecar still has no `--metrics-address` (args read in full).
+- **Flux ordering forbids shipping the fix from here:** `oauth2-proxy dependsOn=['platform-config']`
+  and `apps-wave-0 dependsOn=['platform-config']` — siblings, no ordering. A NetworkPolicy naming ns
+  `jackett` from this tier would apply today and fail on a cold rebuild.
+- **The `== 1` guard is non-vacuous — proven by mutation.** Reverting it to the bare `up` makes the
+  two new scenarios FAIL (`got: nil` at 1h5m and 2h5m) and leaves every other assertion passing;
+  restoring it returns the suite to SUCCESS under pinned promtool v3.14.0.
+- **Gate:** see the next section for the exit code and skips.
+
+## What I did NOT verify
+
+- **The PodMonitor has never been applied.** That it discovers the five pods is reasoned from the
+  measured pod labels and port name, not observed — no target list has been read with it in place,
+  because applying it is an operator action. **Expect `up == 0` on all five when it lands**, which is
+  the documented state, not a regression.
+- **I did not fix the NetworkPolicy, and this PR therefore does not make the gates observable.** It
+  makes them *discovered*, and makes the remaining gap loud instead of silent. The one narrow rule
+  per app (ns monitoring → `auth-mode=forward-auth` → TCP 44180, per
+  `platform/longhorn/networkpolicy-metrics.yaml`) has to be rendered with each app's own policy, in
+  the catalog/renderer that is in flight on another branch. **It needs an owner.**
+- **Whether Prometheus will actually select a PodMonitor in ns `oauth2-proxy` targeting other
+  namespaces.** `namespaceSelector.any: true` plus `podMonitorSelectorNilUsesHelmValues: false` and
+  the fact that cross-namespace ServiceMonitors/PrometheusRules already work here say yes; it is not
+  measured with this object.
+- **No session-lifecycle claim can be made from the current window.** `AuthSuccess` is **0** on both
+  jackett and radarr since the PR #28 restart, so nobody has signed in and `--cookie-refresh=2m30s`
+  has never executed a refresh. The cookie fix is unproven in the one dimension that matters most.
+- **The radarr `/signalr` finding is the coordinator's measurement, not mine** — I reproduced neither
+  the 206/209 sample nor the jackett 0-in-13,612. They are recorded as reported.
+- **L2/L3 conformance still did not run** (`CLUSTER_AVAILABLE` unset). Nothing is certified.
+
+### Round 2 addendum — CI is RED, by design, and the local gate cannot see why
+
+After pushing round 2, `gate / gate (pull_request)` reported **failure after 1m42s** while the local
+toolbox run was rc 3 with nothing failed. Measured cause, singular:
+
+    sensitive paths touched:
+      conformance/tests/l0-static/servicemonitor-ports-resolve/check.sh   (matched conformance/**)
+    approving reviews: (no reviews at all)        author=gitea-admin
+
+`ci/approval-policy.sh` is fail-closed and exits 1 when a PR touches a `sensitive_paths` glob with no
+`APPROVED` review from a non-author — "a change to the gate cannot be gated by the gate". The single
+`KNOWN_EXTERNAL_MONITORS` entry is what trips it. **Nothing in L0/L1 failed.**
+
+**The local gate structurally cannot catch this, and said `approval-policy: green (N/A)`.** The stage
+short-circuits to N/A when `GITHUB_EVENT_NAME` is unset, which it always is off-CI. So for any branch
+touching a sensitive path, a clean local gate is a FALSE REASSURANCE — the PR run is the only honest
+signal. That is a general trap, not specific to this change, and it is worth knowing before someone
+concludes a red PR means broken code.
+
+To clear: a non-author approving review, then **re-run the gate** — approval does not re-trigger the
+status check, so the required context stays red until the job runs again. `required_approvals=1`
+already applied to every PR here, so the human was needed regardless; the conformance entry only
+turns the tick red while that is pending.
+
+The alternative is to drop the register entry, which means dropping `podmonitor.yaml` with it (the
+check fails without the declaration). I did not take it: the entry is the "declare the limit"
+mechanism the check explicitly asks for, it is shrink-only so it cannot rot quietly, and
+`flux-podmonitor.yaml` is the same case already in the register. But it is a real trade and the
+reviewer should make it knowingly.
@@ -0,0 +1,378 @@
+---
+date: 2026-09-30
+branch: fix/oauth2-proxy-cookie-collision-fleet
+title: The cookie collision was never Crafty's — it was the renderer's, and it shipped to five apps
+---
+
+## What changed
+
+Follow-up #3 of `decisions/2026-09-29-crafty-oauth2-proxy-signin.md`: *"Any further per-app
+reverse-proxy sidecar needs a unique `--cookie-name` too. The collision is a property of the
+pattern, not of Crafty."*
+
+It is a property of the pattern, and this branch found where the pattern actually lives.
+
+**Read the scope claim carefully before the detail.** This branch fixes THREE defects in the
+forward-auth sidecar contract, all in one function, all landing on all five apps at once:
+
+| Defect | Fix | Measured evidence |
+|---|---|---|
+| Six independent instances, seven different cookie-secrets, **one** cookie name on **one** zone-wide domain — so they overwrite and `ClearSession`-delete each other's sessions | `--cookie-name=_oauth2_proxy_<app>` + host-scoped `--cookie-domain` | 7/7 distinct secrets; 6 instances on the default name; radarr logs the eviction fingerprint 8× in 12h |
+| The session dies **5 minutes** after sign-in — `--cookie-refresh` defaults to 0 = *disabled*, so the refresh token the proxy already holds is never spent | `--cookie-refresh=2m30s` (half the lifetime) | access-token lifetime = **300.000s exactly**, `refresh_token:true`, flag absent |
+| **No** per-app sidecar serves `/metrics` at all — without `--metrics-address` the listener never starts | `--metrics-address=0.0.0.0:44180` + a `metrics` container port | `ports` was `['http']` on all six |
+
+Rejection *volume* is dominated by the second defect, not the first: ~95% of radarr's rejections
+carry no cookie at all. **Do not merge this expecting one number to drop** — three mechanisms were
+tangled together, and separating them is most of the value here.
+
+**Boundary note.** My brief originally reserved `apps/wave-0/jackett.yaml` for a concurrent branch.
+That split is not achievable and has been revoked: the file is generated, the fix is one renderer
+edit, and `render-golden` requires all five goldens regenerated in the *same commit*. I own all
+five. The proof I ran before touching jackett is under "What I verified".
+
+**The defect was not in the four manifests it was reported in.** `apps/wave-0/pihole.yaml`,
+`apps/wave-0/qbittorrent.yaml`, `apps/wave-1/radarr.yaml`, `apps/wave-1/sonarr.yaml` — and
+`apps/wave-0/jackett.yaml` — are **generated goldens**. One function emits all five sidecars:
+`catalog/render.py::_oauth2_sidecar_container()`, which hardcoded `--cookie-domain=.{ZONE}` and
+emitted no `--cookie-name` at all. Hand-editing a golden is not a fix twice over: `ci/render-golden.sh`
+FAILS on the drift (provoked, not assumed — see below), and the next re-render silently reverts it.
+
+So the fix is in the generator, and both flags are now **derived** from the ServiceDefinition:
+
+    --cookie-name=_oauth2_proxy_{s['name']}     unique per service, by construction
+    --cookie-domain={host}                      the app's own route host, not the zone
+
+`--whitelist-domain` deliberately stays zone-wide — it governs permitted post-login *redirect*
+targets, not cookie scope. `catalog/render.py --selftest` gained four assertions, including one
+that renders a **second** synthetic forward-auth service and requires the two cookie names to
+differ, so a future refactor to a constant fails in the renderer's own test.
+
+Re-rendering with the canonical command regenerated **five** goldens, not four. Jackett is the
+fifth, and this is the one thing in this branch the PM brief did not anticipate — see
+"the jackett collision" below. The three `conformance/tests/l1-rendering/*/golden/` fixtures are
+expected-output copies of the same renderer, so they are regenerated too; that is not collateral
+damage, it is the L1 stage doing its job (it FAILED on the first gate run and named the drift).
+
+**The new gate check** is `conformance/tests/l0-static/oauth2-proxy-cookie-names-distinct/`,
+declaring `conformance.home.k3s/scope: whole-repo` so `ci/l0.sh`'s `--include-scope` union runs it
+on **every** gate invocation. No bespoke `ci/l0.sh` stage was added — `whole-repo-invariants-wired`
+fails the gate if anyone adds one, and the scope annotation is the supported mechanism.
+
+It inventories every oauth2-proxy instance the repo authors (containers at any depth, plus the
+shared proxy's HelmRelease values) and resolves each one's **effective** cookie name — its declared
+name, **or oauth2-proxy's built-in `_oauth2_proxy` when it declares none**. That single modelling
+decision is the whole check: an omitted flag is not an absence, it is a *claim on a name*, and
+treating it as an absence is precisely why five reviewers read these manifests and saw nothing.
+Then: (1) no two effective names collide; (2) no name is another's name + `_csrf`, because that is
+how oauth2-proxy derives the CSRF cookie and two such session names look distinct while their CSRF
+cookies collide; (3) the default name has exactly one owner, and it is the one registered in
+`SHARED_DEFAULT_COOKIE_OWNER`; (4) every non-registered instance is host-scoped; (5)
+`catalog/render.py` derives rather than lists; (6) the **shipped** renderer is proven end-to-end on
+two synthetic ServiceDefinitions run as a subprocess.
+
+**On the shared proxy — "distinct" or "explicitly declared"?** The brief asked me to decide, and
+the answer is *neither, quite*. The shared forward-auth proxy at `platform/oauth2-proxy/` **must
+keep** `_oauth2_proxy`: it is the single session every forward-auth-gated surface without its own
+sidecar rides on, so renaming it signs all of them out at once. Requiring *every* instance to
+declare a name would therefore need an exemption, and an exemption is a hole: delete the shared
+proxy and whichever app next omitted the flag would silently inherit the default with the rule
+still satisfied. So the assertion is plain **distinctness over effective names**, with the default
+treated as a name like any other — held by **exactly one owner, named in a shrink-only register
+with its reason**. The shared proxy needs no exemption from the rule; it is simply the registered
+holder of that one name. A stale register entry FAILS (the exemption would outlive its object) and
+an unregistered holder FAILS (no silent heirs). Both were provoked and are quoted below.
+
+**On host-scoped `--cookie-domain` — overreach, or not?** Asserted, and here is the argument both
+ways. *Against:* distinct names alone close the eviction bug completely — two differently-named
+cookies do not overwrite each other however wide their domains are — so this assertion fixes
+nothing (1) has not already fixed, and a check that fails a legitimate configuration is a check
+someone eventually disables. *For, and this wins:* a zone-wide cookie is **transmitted by the
+browser to every host in the zone**. qBittorrent's session token was being sent to pihole, radarr,
+sonarr and every other name under the zone, including hosts behind no proxy at all — a
+session-credential disclosure across trust boundaries that nothing needs. And it is what makes a
+*future* name collision catastrophic rather than contained: with host-scoped cookies, two
+instances that accidentally share a name cannot see each other's cookie at all. Defence in depth
+on the exact axis that already failed once. The escape hatch is the same register, by name and
+with a reason, rather than a loosened rule.
+
+### The jackett boundary, and why it could not hold
+
+`apps/wave-0/jackett.yaml` carries the identical defect and was initially assigned to a concurrent
+branch, with an instruction to me not to touch it. **That separation is not achievable** — the
+ownership split was drawn across a generated file — and I did not discover this by reasoning about
+it, I provoked it. With the renderer fixed and jackett's golden left at its `main` content,
+`./ci/render-golden.sh` exits **1**:
+
+    [fail] GOLDEN DRIFT: apps/wave-0/jackett.yaml does not match the canonical render:
+    -        - --cookie-domain=.${BASE_DOMAIN}
+    +        - --cookie-name=_oauth2_proxy_jackett
+    +        - --cookie-domain=jackett.${BASE_DOMAIN}
+    [fail] render==golden check FAILED
+
+Jackett's golden is an *output* of the file that contains the defect. There are exactly three
+options and two of them are worse: hand-edit four goldens and leave the generator broken (fails
+`render-golden`, and reverts on the next render); add a per-service carve-out to the renderer so
+jackett alone keeps defaulting (ships the exact "list, don't derive" anti-pattern this check
+exists to forbid, and leaves jackett broken *by design*); or fix the generator and regenerate all
+five. I took the third. Jackett's golden here is **machine output of a one-line generator change**,
+not an independent edit of Mira's file, and it is byte-identical to what her branch must produce.
+
+**Consequence:** the jackett branch has no cookie work left in it — this commit fixes jackett
+identically and by the same mechanism, and both branches independently arrived at the *same two
+values* (`_oauth2_proxy_jackett`, `jackett.${BASE_DOMAIN}`), which is a good sign for the
+diagnosis. Its hand-edit of the rendered file will fail `render-golden` on its own terms, so it
+should be closed rather than rebased. **What must not be lost with it** is its evidence: the
+interleaved radarr/jackett log showing `removing session` deleting a clean session 26s after login,
+and the `expires − created = 300.000s` measurement that produced `--cookie-refresh` here. Both are
+carried into this entry and into the renderer's comments; the branch itself is now redundant, the
+work behind it is not.
+
+**The general lesson, which is worth more than this bug:** an ownership split across
+`apps/wave-*/*.yaml` cannot hold, because those files are *outputs*. Two agents assigned two
+rendered apps are assigned the same function and will collide in `render-golden` no matter how
+carefully they stay in their own files. Ownership of a rendered app is ownership of
+`catalog/render.py`, and briefs should say so.
+
+### What this branch deliberately does NOT do
+
+- **`--api-route=^/api/` is NOT added**, and not because of scope discipline alone: it would be
+  **wrong** renderer-wide. It was measured on jackett, and **PiHole does not serve `/api/`** — so
+  a renderer constant applies one app's path shape to five. Doing it properly means declaring API
+  path prefixes per service, and the `auth` block is `additionalProperties: false`, so that is a
+  *schema change*. Separate PR. (Note the flag is **`--api-route`, singular** — verified in the
+  pinned binary's own `--help`. The plural `--api-routes` does not exist in v7.8.1 and the proxy
+  refuses to start with it. That error was propagating through briefs; it stops here.)
+- **`--session-cookie-minimal` is NOT added** and must not be, as a cookie-size fix: it is
+  **mutually exclusive** with `--cookie-refresh` (*"cookie_refresh > 0 requires oauth tokens in
+  sessions"*). Dropping `--pass-access-token` is a least-privilege argument, never a size one.
+- **Metrics are NOT scraped, only served.** The flag and container port are the prerequisite;
+  a container port is not a Service port. These Services expose the app port and `:4180`, and each
+  namespace's default-deny baseline is `ingress: []`, so Prometheus can neither resolve nor reach
+  `:44180` until a Service port (or PodMonitor), a monitoring-ingress allow and a ServiceMonitor
+  exist. That wiring is someone else's follow-up and I have deliberately not faked its start.
+  **Nobody should read this commit as "the sidecars have metrics now."**
+- **The new check was NOT widened to cover cookie-refresh or metrics.** Its subject is cookie
+  *identity*; those two share neither its inputs nor its failure mode. Bundling them would make
+  the check's name a lie and couple unrelated ratchets, so a future metrics decision could only
+  be expressed by weakening a *cookie* check. The cost of that discipline is named honestly in
+  the follow-ups: those two flags are currently **ungated**.
+
+## What I verified
+
+- **The premise, in-cluster, rather than inherited from the handoff** (CLAUDE.md §2b). Read-only,
+  `kubectl --context k3s-prod`, through the pinned toolbox.
+  - **Seven** `*-oauth2-proxy` Secrets carry a `cookie-secret`: crafty, jackett, pihole,
+    qbittorrent, radarr, sonarr, and the shared `oauth2-proxy-credentials`. **All seven values are
+    DISTINCT.** Compared without printing anything: each value was hashed **in-cluster-side script,
+    never echoed**, and only the first 8 hex characters of each `sha256` were emitted — enough to
+    say same-or-different, recovering nothing. Secret *key names* were listed; no `.data` or
+    `.stringData` was printed (CLAUDE.md §6). **So no instance can decrypt another's cookie — it can
+    only overwrite it. The premise holds.**
+  - **The default cookie names really were in use.** Of 8 live oauth2-proxy containers, **6 declare
+    no `--cookie-name` at all** — jackett, pihole, qbittorrent, radarr, sonarr, and both replicas of
+    the shared proxy — so all six emit `_oauth2_proxy` / `_oauth2_proxy_csrf`. All six also carry
+    `--cookie-domain=.home.keithjackson.dev`. **One cookie pair, six independent instances.**
+  - Crafty's live pod (`crafty-647c7d8cc5-5spv9`) already carries
+    `--cookie-name=_oauth2_proxy_crafty` and `--cookie-domain=crafty.home.keithjackson.dev`, so
+    PR #22 is merged and reconciled. It is the only compliant per-app instance on the cluster.
+- **THE COLLISION IS REAL. IT IS NOT ESTABLISHED AS THE CAUSE OF THE REPORTED FAILURES.** These
+  are two different claims and only the first is supported. I was handed counter-evidence —
+  crafty, the app already fixed, showed the *worst* rejection ratio of the set over a 24h Loki
+  window (`crafty 81:1`, `radarr 67:1`, `jackett 39:4`) — and it is right to refuse the causal
+  claim. What I could do was bound the window. Read from **each sidecar's own log** rather than
+  Loki, because `kubectl logs` on a pod that has not restarted cannot reach a pod older than
+  itself, so a window inside the pod's age is inside the fixed spec's lifetime *by construction*:
+
+      namespace      age(h) rst  reject success    ratio  cleared  badsig
+      crafty           13.9   0       3       0      n/a        0       0   <- FIXED spec, whole life
+      radarr          422.5   0     155       1  155.0:1        8       8
+      jackett         422.6   0      39       4    9.8:1        1       1
+      pihole          395.7   0       0       0      0:0        0       0
+      qbittorrent     404.1   0       0       0      0:0        0       0
+      sonarr          422.5   0       0       0      0:0        0       0
+      oauth2-proxy    422.6   0       0       0      0:0        0       0   (both replicas)
+
+  `reject` = `No valid authentication in request`; `success` = `AuthSuccess`;
+  `cleared` = `removing session` (oauth2-proxy's `ClearSession`); `badsig` = `cookie signature
+  not valid`. Window `--since=24h`, which exceeds every pod's age only for crafty — whose pod is
+  13.9h old with **0 restarts**, so these are its complete post-fix figures.
+
+  Then I read crafty's four auth-relevant lines rather than counting them, because at n=3 the
+  content is the evidence:
+
+      03:57:20  Cookie settings: name:_oauth2_proxy_crafty secure(https):true httponly:true
+                expiry:168h0m0s domains:crafty.home.keithjackson.dev path:/
+      04:02:55  No valid authentication in request. Initiating login.
+      15:22:39  No valid authentication in request. Initiating login.
+      15:22:44  No valid authentication in request. Initiating login.
+
+  6743 log lines; those are all four. **Zero** `removing session`, **zero** `cookie signature
+  not valid`, zero anything else.
+
+  Four conclusions, kept separate:
+  1. **`crafty 81:1` is a window artefact. It is not evidence against the fix and must not be
+     cited as such.** The fixed pod emitted **3** rejections in 13h 52m; the 81 came from a 24h
+     window straddling it, i.e. from the pre-fix process.
+  2. **The eviction FINGERPRINT is gone where the fix is deployed.** `cookie signature not
+     valid` + `removing session` is what a clobbered instance logs — a cookie *is* present,
+     written with a key this instance does not hold, so it cannot decrypt it and then
+     `ClearSession` expires it on its own `--cookie-domain`. Fixed crafty: **0 and 0** across
+     its whole life. Unfixed radarr: **8 and 8** in 12h. Unfixed jackett: 1 and 1. That is a
+     real, directional signal and it is the strongest evidence available that the fix works.
+  3. **It is still not a demonstrated successful sign-in.** Crafty logged **0 `AuthSuccess`** —
+     so what is shown is *absence of the failure mode*, not a session observed working
+     end-to-end. Those three rejections are consistent with benign first-hits (one client, two
+     of them 5s apart, i.e. a page plus a `/ws` reconnect). I am comfortable saying the fix
+     works; I am not willing to say it has been watched working, and the distinction is the
+     whole reason this section exists.
+  4. **The collision is a MINORITY of the rejection volume.** radarr: 8 of 155 (5%) carry the
+     fingerprint; ~95% carry **no cookie at all**. Those are the 5-minute session expiry and
+     `302`-on-XHR — the first of which this branch now also fixes, the second of which it
+     deliberately does not.
+
+  Two cautions about the instrument. `reject` ranks **chattiness, not breakage**: radarr's are
+  `/api/v3/command`, `/api/v3/queue/details` and `/signalr/negotiate` polls, so a busy *arr
+  inflates the count with nobody inconvenienced. And that same polling is the *amplifier* that
+  made this intermittent — every poll carrying an undecryptable cookie calls `ClearSession`
+  again on the zone-wide domain, so **a left-open Radarr tab deleted Jackett's cookie every few
+  seconds, from another namespace, with no user action at all.** That is why signing in again
+  never helped for longer than the gap between two polls. It is now a comment in the renderer,
+  because it is the part of this bug that no one would reconstruct from the flags alone.
+- **The four target files are generated, not authored.** Rendering `catalog/` with the canonical
+  command into a temp dir and diffing against `apps/` reported all 10 onboarded goldens
+  byte-identical before any change. Hand-editing one is drift by definition.
+- **`render-golden` FAILS if jackett is excluded** — provoked (output quoted above), rc **1**.
+  Restoring the regenerated golden returns it to `all 10 onboarded service golden(s) reproduce`,
+  rc **0**.
+- **Every flag verified against the PINNED v7.8.1 binary**, not against docs, memory, or the
+  shared proxy (which runs **v7.15.3** and therefore could not vouch for v7.8.1):
+  `podman run --rm quay.io/oauth2-proxy/oauth2-proxy:v7.8.1 --help`.
+  - `--cookie-name string ... (default "_oauth2_proxy")` — **this confirms the constant my whole
+    check is built on, from the binary itself.** The check's central claim, that an omitted flag
+    is a *claim on a name* rather than an absence, now rests on the implementation's own help
+    text rather than on my reading of its source.
+  - `--cookie-refresh duration  refresh the cookie after this duration; 0 to disable` — confirms
+    the default really is disabled.
+  - `--metrics-address string` — present in v7.8.1, so the new flag cannot crashloop the pods.
+  - `--api-route strings` — **singular**, confirming the brief's plural was wrong.
+  - `--session-cookie-minimal` — exists, and is the flag that conflicts with `--cookie-refresh`.
+- **44180 is the platform's existing convention, measured, not chosen.** The shared proxy already
+  runs `--metrics-address=0.0.0.0:44180` with a `metrics` container port, a `metrics:44180`
+  Service port, a ServiceMonitor whose endpoint is `metrics`, and **2 ready endpoint addresses** —
+  i.e. it demonstrably works on this platform.
+- **`catalog/render.py --selftest`: PASS**, now including the two-service distinctness assertion,
+  a `--cookie-refresh` assertion that PARSES the duration and requires `0 < refresh < 300s` (so
+  `2m30s` -> `150s` -> `2.5m` all stay green but a silent widening past the token lifetime does
+  not), a `--session-cookie-minimal` mutual-exclusion guard, and both metrics assertions.
+- **The check is selected by the whole-repo union.** `conformance/run.sh --layer l0 --requirements
+  r-nosuchreq-99 --include-scope whole-repo --print-selection` lists
+  `tests/l0-static/oauth2-proxy-cookie-names-distinct` among 20 tests — i.e. it runs even when the
+  touched-requirement filter matches nothing, which is the blind spot it was scoped for.
+- **The check was tested in BOTH directions**, against throwaway trees outside the repo, never by
+  mutating the branch:
+  - **Compliant tree (this branch): rc 0, 39 assertions, 7 instances audited.**
+  - **Pre-fix tree** (`git archive HEAD` + this test dropped in): **rc 1, 26 failures**, naming all
+    five violators, the 6-way `_oauth2_proxy` collision, the zone-wide domains, the renderer
+    emitting no `--cookie-name`, and both synthetic fixtures rendering `None`.
+  - **Targeted mutations on a compliant copy**, one assertion class each, all fired and nothing
+    else did: duplicate name (sonarr←radarr) → 1 failure; CSRF derivation
+    (`_oauth2_proxy_sonarr_csrf` as a session name) → fired; zone-wide `--cookie-domain` on pihole
+    → 2 failures; crafty's `--cookie-name` deleted → unregistered-default-heir + 2-way collision;
+    shared proxy renamed → **stale register entry** failure.
+  - A **false positive was found and fixed** by this testing: `platform/oauth2-proxy/forward-auth-cec.yaml`
+    contains `serviceName: oauth2-proxy/oauth2-proxy:80` — an Envoy `namespace/service:port`
+    reference, not an image — which the first inventory scan counted as an unaudited instance. The
+    scan is now anchored on `image:`, and the near-miss is documented in the check.
+- **Gate:** toolbox `./ci/gate.sh` — exit code and stage list in the next section, because it is
+  not a pass and that belongs there.
+
+## What I did NOT verify
+
+- **Local gate: rc 3 (UNDETERMINED), not 0.** Green: `spec-immutability`, `decisions-log`, `lint`,
+  `l0`, `render-golden`, `l1`, `approval-policy`. **Skipped:** `gitea-settings` (no
+  `SETTINGS_API_TOKEN` locally) and **L2/L3 conformance** (`CLUSTER_AVAILABLE` unset). Nothing
+  FAILED; nothing runtime was certified either. **rc 3 is UNDETERMINED, never green.** Touched
+  subset `R-GITOPS-02 R-IDP-02 R-IDP-06 R-NET-01 R-NET-05 R-NET-06 R-NET-07 R-PLAT-02 R-TEST-01
+  R-TEST-02 R-TEST-03 R-TEST-04` — **UNVERIFIED** at L2/L3, all twelve.
+- **`spec-immutability` was skipping because of a BUG IN MY TOOLING, not the environment.** The
+  worktree ENGINE shim in circulation mounts only the main checkout's `.git`, and `dev.sh`
+  computes the spec path from its own `REPO_ROOT` — which in a worktree is
+  `.claude/worktrees/agent-*/`, so `../home-network-k3s` does not resolve and it adds no spec
+  mount at all. The gate then reported the skip, and the skip read as "the one tolerated
+  environmental skip" while actually being a path bug being waved through. Adding
+  `-v /var/home/keith/source/home-network-k3s:/home-network-k3s:ro -e SPEC_REPO=/home-network-k3s`
+  makes **R-TEST-06 actually run and pass**. Worth putting in every worktree agent brief; the
+  two-mount shim in CLAUDE.md §3 should say so.
+- **CI: `gate / gate (pull_request)` reports FAILURE on `75a4029`, for exactly ONE reason, and it
+  is not this change.** Read from the job log rather than the tick:
+
+      [fail] approval-policy: a sensitive change lacks a non-author approval -> blocks merge
+      [fail] MERGE GATE: BLOCKED
+
+  Every substantive stage is **green in CI** — `gitea-settings: pass` (CI has the token my local
+  run does not), `decisions-log`, `lint`, `l0`, `render-golden`, `l1`. Skipped: `spec-immutability`
+  (the workflow does not check the spec repo out — the documented tolerated skip) and
+  `conformance` (declared). So CI is a **stricter, more complete** run than my local one on every
+  axis except the spec repo, and it agrees.
+  `infra/gitea/approval-policy.json` lists `conformance/**` as sensitive — *"the tests and the
+  harness; changing them changes what green MEANS"* — so this PR is **fail-closed pending a
+  non-author approving review**, which is CLAUDE.md §1 working, not a defect to route around.
+  **The red tick is the policy, and it clears when a human reviews it.** Note also that
+  `gate / conformance (pull_request)` reports `success` with description **"Has been skipped"** —
+  the precise trap CLAUDE.md §2b names; it certifies nothing.
+- **A SUCCESSFUL SIGN-IN, ANYWHERE.** The strongest post-fix evidence is the *absence* of the
+  eviction fingerprint on crafty (0 `removing session`, 0 `cookie signature not valid`, across
+  13h 52m) against radarr's 8-and-8. That is directional and I believe it. But crafty logged
+  **0 `AuthSuccess`** in the same window, so no session has been observed working end to end on a
+  fixed instance. Whoever closes this out should re-measure crafty **after real sign-in traffic**.
+- **No browser round-trip by me at all**, for any of the three defects. Reproducing the collision
+  needs two gated apps visited in one session. The concurrent jackett branch does carry an observed
+  log-level reproduction (interleaved radarr/jackett `removing session` deleting a clean session
+  26s after login) — real evidence for the MECHANISM, and it is theirs, not mine.
+- **`--cookie-refresh=2m30s` has never been exercised.** The value is derived from a measured
+  300.000s access-token lifetime and is asserted to sit inside `(0, 300s)`, but no session has been
+  held past the five-minute cliff to watch a refresh succeed. If Keycloak's refresh-token or SSO
+  idle timeouts are shorter than expected, the symptom moves rather than disappears. **This is the
+  fix most likely to need a second pass**, and the first thing to watch after merge.
+- **The `metrics` port has never been fetched.** The listener is asserted to start by flag
+  presence only; nobody has curled `:44180` on a rendered pod. And it is unreachable by design
+  until the Service/policy/ServiceMonitor wiring lands — see above.
+- **Nothing was applied.** No `kubectl apply/edit/patch`, no `flux reconcile`. Cluster access was
+  read-only throughout; the one container I ran was the oauth2-proxy image locally under podman to
+  read its `--help`. The running pods are still the pre-fix spec. **Merging is what deploys**, and
+  it will roll five Deployments; every user's existing session on those five apps is invalidated by
+  the cookie rename and they will re-authenticate once. Radarr/sonarr/qbittorrent use
+  `strategy: Recreate` on RWO volumes, so there is a brief gap per app. Pick the moment — and
+  note this now changes **three** flags plus a port on all five, so a bad `--cookie-refresh` value
+  would surface as five apps misbehaving, not one.
+- **The live flags are not re-measured by the check.** git is not the cluster. I measured the live
+  pods by hand, once; nothing in the gate re-reads them, so hand-applied drift would be invisible
+  to it. A live-state assertion is L2 and is not written.
+- **`--cookie-refresh` and `--metrics-address` are UNGATED, and I found out why.** Nothing in
+  `ci/` or `.gitea/` runs `catalog/render.py --selftest` — I grepped for it. So the twenty-odd
+  assertions in that selftest, including the six I just added, are dev-time only: a future edit
+  that drops either flag *and* re-renders the goldens would pass the entire merge gate.
+  `render-golden` catches accidental drift (golden vs renderer) but not a deliberate removal of
+  both together, and my L0 check deliberately does not cover these two flags. **The repo already
+  has the precedent for the fix and the one-line shape of it:**
+  `conformance/tests/l0-static/auditability-proofs-producible/check.sh:291` drives
+  `conformance/lib/<tool> --selftest` as a subprocess for exactly this reason — its own header
+  says it does so "because nothing [else] drives them". Adding `catalog/render.py` to that list,
+  or giving the renderer its own scope-declaring L0 test, closes it. I did not do it here because
+  it widens an already review-blocked PR into a third subject; it is the highest-value follow-up
+  in this entry.
+- **Log-derived counts are a weak instrument and I did not strengthen them.** `kubectl logs
+  --since=` returns what the kubelet still retains, so an old, busy container's window may be
+  truncated; and `No valid authentication in request` counts unauthenticated first hits and XHR
+  polls indiscriminately. The `cookie signature not valid` / `removing session` pair is the only
+  collision-specific signal in there, and it is the one I leaned on. None of this is gated —
+  there is no alert on it and no L2 check reads it.
+- **The cookie-secrets' distinctness is not gated either.** It was measured out of band and is
+  asserted nowhere. If they were ever made identical the collision would become harmless — but
+  sharing one session-encryption key across trust zones is a worse idea than the bug it would mask,
+  so the check does not offer it as an alternative and does not look for it.
+- **The shared proxy's own configuration is unchanged and untested by me.** I did not touch
+  `platform/oauth2-proxy/**` (owned by another agent this cycle); the check only reads it.
+- **Nothing about group authorization, CSRF behaviour, or session contents.** Names and domains only.
+- **`vars.CLUSTER_AVAILABLE` is still unset in CI**, so the PR's green tick — if it goes green —
+  proves the L0/L1 offline subset and nothing about runtime.
@@ -0,0 +1,174 @@
+---
+date: 2026-09-30
+branch: feat/plex-quicksync
+title: Plex opts into `hardware.gpu: intel-quicksync` — the device mount is GitOps, the Plex-side toggle is a documented manual step, and hardware transcoding has NOT been observed working
+---
+
+## What changed
+
+Three things, and the third is the point of the branch.
+
+1. **`catalog/plex.yaml`** — added the four lines the whole feature reduces to:
+
+   ```yaml
+   hardware:
+     gpu: intel-quicksync
+   ```
+
+   placed between `env` and `exposure` to match the schema's own property order, with ~48 lines of
+   comment in the file's established style. The comment records: the field is **intent, not
+   mechanism** (a later move to the Intel device plugin is a `render.py` change and this block does
+   not move); that `/dev/dri/renderD128` is present on **all three** prod nodes with `i915` bound;
+   **why a `supplementalGroups` entry would be inert** (the image's cont-init hook does the group
+   work inside the container and `s6-setuidgid` replaces the process group list from `/etc/group`);
+   why `DirectoryOrCreate` is the load-bearing safe failure mode (`apps-wave-0` is `wait: true` with
+   `apps-wave-1` depending on it, so a failed mount would stall the whole app tier for a feature
+   that is not load-bearing); and that the Plex-side toggle is a **separate manual operator step**,
+   pointing at the new runbook.
+
+2. **`apps/wave-0/plex.yaml`** — re-rendered. Six added lines, nothing else, in the whole `apps/`
+   tree:
+
+   ```
+           - name: dri
+             mountPath: /dev/dri
+         - name: dri
+           hostPath:
+             path: /dev/dri
+             type: DirectoryOrCreate
+   ```
+
+3. **`runbooks/plex-quicksync.md`** (new), registered in `runbooks/kustomization.yaml`'s
+   `configMapGenerator` as `- plex-quicksync=plex-quicksync.md` and indexed in `runbooks/README.md`
+   under "Operator procedures". **The file alone would have been a 404** — the pod serves a
+   ConfigMap, not the git tree, and `conformance/tests/l0-static/alert-runbooks-resolvable` exists
+   precisely because "the file is in git" and "the URL resolves" are different claims. The page
+   covers: the GitOps/manual boundary as a table; the toggle (Settings → Transcoder → "Use hardware
+   acceleration when available", preference key `HardwareAcceleratedCodecs`); the real neighbouring
+   keys and the widely-cited one that **does not exist**; that the only proof is `Transcode (hw)` on
+   the Plex dashboard **and why no alert is possible**; the device-cgroup failure to expect first
+   with a concrete open() diagnostic and the Intel-device-plugin fallback; and the floating-`:latest`
+   caveat with a digest check.
+
+**Stacked on PR #27 (`feat/sd-hardware-capability`), which is open and unmerged.** This branch was
+cut from it rather than from `main`, so its one commit (`d9baa2a`, the schema + renderer + contract)
+is in this PR's range. That is deliberate: merging #27 is not an agent's action (CLAUDE.md §1), and
+duplicating its changes here would have been worse than showing them. **This must merge after #27.**
+
+Not done, deliberately: no `--emit-waves` (it writes into `clusters/base/`, a sensitive path that
+would force a human reviewer, and it is a no-op unless the *set* of files under `apps/` changes — it
+does not here). No `podSecurityContext`. No `nodeSelector`. No alert rule. No `kubectl apply`.
+
+### Two follow-ups left unclaimed, with the reasoning
+
+- **Transcode scratch is on the wrong medium, and it is invisible.** The image sets
+  `TranscoderTempDirectory=/transcode`, which is **not a volume** — it is the container's writable
+  overlay, i.e. the node's `/data`. That is NVMe on prod-2 but **SATA on prod-1 and prod-3**, so
+  transcode scratch performance changes silently on reschedule, with nothing reporting it. Both
+  candidate fixes need a **schema** change and so neither belongs in this PR: pin Plex to prod-2
+  (needs a `nodeSelector` field the schema does not have), or give `/transcode` a `medium: Memory`
+  emptyDir and transcode in RAM (~31 GiB per node; needs an emptyDir field the schema does not
+  have). **The RAM option is the better one** — it removes the node asymmetry outright instead of
+  working around it — which is exactly why it deserves its own PR and its own review rather than
+  being smuggled in behind a four-line catalog change. Documented in the runbook's "Known gaps".
+- **GuC/HuC firmware is not loaded** (`i915.enable_guc=0`). The cost is low-power H.264 **encode**
+  quality only; decode and this feature are unaffected. Changing it needs a kernel parameter plus a
+  **serialized rolling reboot** — an operator action under CLAUDE.md §5, on a cluster with a
+  documented history of that going badly (a prod node hung on a warm reboot on 2026-09-03). Kept
+  well clear of this PR.
+
+### Rebase note (2026-10-01, Ruth Okonkwo) — resolution only, no redesign
+
+Rebased onto `main` @ `08deb20` (was cut from `0fbf544`), restacked on the rebased
+`feat/sd-hardware-capability` (PR #27) so the stack stays coherent — this branch still carries
+exactly PR #27's commit plus its own, with no duplicate of anything already in `main`. The branch
+ref was force-updated with `--force-with-lease`, so PR #32 is the same PR with a new head.
+
+**The `catalog/render.py` conflict disappeared** once PR #27 was rebased; it was never this
+branch's own conflict. **One conflict remained, in `runbooks/README.md`, purely textual.** PR #30
+(`fix/jackett-oauth2-proxy-session`, merged as `08deb20`) appended an `oauth2-proxy-session` index
+entry under *Operator procedures* at the same point this branch appended its `plex-quicksync` entry.
+**Both entries were kept, in full.** Ordering follows this file's existing convention, which is
+grouped/narrative rather than alphabetical (the *Nodes* section runs down/io/memory/disk/cpu):
+`plex-quicksync` sits next to `plex-claim` as the other Plex page, leaving `oauth2-proxy-session`
+adjacent to `crafty-admin-credential`, which cross-references the same sidecar gate. The change to
+`runbooks/README.md` is an 8-line pure insertion, so PR #30's entry is untouched.
+`runbooks/kustomization.yaml` auto-merged and **both** `configMapGenerator` entries are present.
+
+Re-verified after the resolution, on the rebased stack: `--selftest` PASS (rc 0); a full re-render
+reproduces every committed golden byte-identically; and the **only** `apps/` diff against the new
+`main` is `apps/wave-0/plex.yaml`, confined to the expected six lines (the `dri` mount and its
+`DirectoryOrCreate` hostPath volume). No other service's golden moved — PR #28 added ~176 lines of
+oauth2-proxy sidecar rendering, and this confirms by measurement rather than assumption that Plex
+(`auth: native`, no sidecar) is unaffected by it.
+
+## What I verified
+
+- **The render is exactly the intended change, and nothing else moved.** After
+  `python3 catalog/render.py --all catalog -o apps --wave-layout --no-image-automation` (exit 0,
+  silent): `git diff --stat apps/` = `apps/wave-0/plex.yaml | 6 ++++++`, and
+  `git status --porcelain apps/` = a single ` M apps/wave-0/plex.yaml`. Read line by line, the diff
+  is the `dri` volumeMount and the `dri` hostPath volume, and nothing more. **Specifically checked:**
+  the same invocation also rewrites `apps/wave-1/keycloak.yaml` and `apps/wave-0/vm-demo.yaml`, which
+  are git-tracked but excluded from the golden comparison by `RENDER_GOLDEN_NOT_ONBOARDED` in
+  `ci/render-golden.sh` — a spurious diff there would have landed **ungated**. Neither file shows a
+  diff; whole-repo `git status --porcelain` lists only `apps/wave-0/plex.yaml` and `catalog/plex.yaml`
+  plus my new files.
+- **The rendered output matches what the contract promises** — `schemas/renderer-contract.md`
+  `### hardware` says exactly two things (a `/dev/dri` hostPath with `type: DirectoryOrCreate` and a
+  read-write mount on the app container) and exactly two things appeared. No `securityContext`, no
+  `nodeSelector`, no resources in the diff.
+- **The runbook is projected into the server, not merely committed.** `runbooks/kustomization.yaml`
+  carries `- plex-quicksync=plex-quicksync.md`, whose shape matches the generator-entry regex the
+  L0 check parses; the extensionless key is what makes `/runbooks/plex-quicksync` a plain static
+  lookup. Also indexed in `runbooks/README.md`.
+- **The runbook clears the L0 quality floor** that `alert-runbooks-resolvable` applies to
+  alert-linked pages, even though nothing links to this one (it is advisory for an unreferenced
+  page): all four required sections (`## What this means here`, `## First commands`,
+  `## Likely causes, most likely first`, `## When to escalate or stop`), well over the 1500-byte
+  floor, `dev.sh` used throughout, and `--context k3s-prod` on every command.
+- **The gate.** See the run recorded in the PR description for the per-stage table and the real rc.
+- **No secret was printed.** Every `Preferences.xml` read in the runbook greps one named key, and the
+  page states explicitly why: the file holds `PlexOnlineToken`. Nothing was `sops -d`'d.
+
+## What I did NOT verify
+
+**Hardware transcoding has NOT been observed working.** Nothing in this branch is evidence that it
+does. A green gate here proves the manifest renders and the golden matches — it proves nothing about
+runtime behaviour, and `vars.CLUSTER_AVAILABLE` is unset so L2/L3 conformance did not run at all.
+
+- **The device-cgroup question is OPEN, and it is the largest remaining technical risk.** Whether
+  containerd on this k3s build permits a **non-privileged** pod to `open()` a bind-mounted character
+  device is **UNVERIFIED**. It can only be settled by deploying and trying. The container inits as
+  uid 0, which settles *file* permissions, but uid 0 does **not** bypass the cgroup device gate — and
+  the two failures look identical from the outside. If it is refused, the fallback is the Intel
+  Device Plugin (`gpu.intel.com/i915`), which is a `catalog/render.py` change plus a re-render with
+  **no ServiceDefinition edit**; `privileged: true` is not available (Kyverno
+  `restrict-host-namespaces` is Enforce). The runbook has the diagnostic sequence.
+- **The Plex-side toggle was still unset at the time of writing.** `HardwareAcceleratedCodecs` was
+  **absent** from `Preferences.xml` — the toggle has never been enabled on this server. Until an
+  operator ticks it in the UI, this branch changes the pod spec and **nothing about transcoding**.
+  That is the whole reason the runbook exists.
+- **No metric, so no alert, so no automated detection — permanently.** `catalog/plex.yaml` declares
+  no `metricsEndpoint`, so Plex renders no ServiceMonitor; there is no transcode-mode signal for a
+  rule to evaluate. A silent degradation to software transcoding (a node losing its iGPU, the
+  `DirectoryOrCreate` path) will **not** be reported by anything. I deliberately did not add an
+  alert: with no metric behind it, it could never fire, and this repo has been burned at scale by
+  exactly that. Closing the gap honestly means an exporter first.
+- **I did not re-derive this session's node or in-container measurements.** The three-node
+  `renderD128`/`i915`/GID-992 facts, the cont-init hook behaviour, the `s6-setuidgid` group
+  replacement, the `PLEX_PREFERENCE*` absence, the six `setPref` keys and the
+  `HardwareAcceleratedCodecs` string were established earlier in this session and are cited, not
+  re-measured, by me. Per CLAUDE.md §2b, treat this file as a record of belief and re-measure against
+  the cluster before relying on any of it.
+- **The image is a floating tag.** Everything known about the hook came from digest
+  `sha256:e0ab27395614…` (`plexmediaserver 1.43.4.10903`), while the Deployment runs
+  `plexinc/pms-docker:latest` with `imagePullPolicy: Always`. A silent pull can invalidate the
+  "no `supplementalGroups` needed" conclusion **with no commit in this repo**. Not guarded against —
+  the runbook tells the reader to re-check the hook after an image bump, which is documentation, not
+  enforcement.
+- **I did not verify the live Gitea gate state** (branch protection, required contexts, runner
+  online) for this PR. Per §2b that must be queried, not read out of `CLAUDE.md`'s table; I did not
+  query it.
+- **`/runbooks/plex-quicksync` was not fetched.** The L0 check asserts every static precondition for
+  an HTTP 200; it does not make the request, and I did not either.
@@ -0,0 +1,204 @@
+---
+date: 2026-09-30
+branch: feat/sd-hardware-capability
+title: ServiceDefinition gains a `hardware` capability field — `gpu: intel-quicksync` mounts /dev/dri and nothing else; the enum is what keeps the mechanism swappable
+---
+
+## What changed
+
+A new optional top-level ServiceDefinition property, `hardware`, with one optional member,
+`gpu`, whose enum today has exactly one value: `intel-quicksync`. **No catalog entry uses it.**
+This PR delivers the capability; consuming it (Plex) is separate work on top.
+
+- `schemas/servicedefinition.schema.json` — adds `hardware` (object, `additionalProperties: false`,
+  `minProperties: 1`) with `gpu` (enum `["intel-quicksync"]`). Schema version bumped **1.7.0 →
+  1.8.0**, changelog entry prepended to the schema `description` per that file's own convention.
+- `catalog/render.py` — two module-level constants (`DRI_VOLUME_NAME`, `DRI_DEVICE_DIR`) and the
+  emission in `r_deployment`.
+- `schemas/renderer-contract.md` — new `### hardware` field-mapping section, plus cross-field
+  invariant **8** in that file's numbered list.
+- `schemas/README.md` — its `**Schema version:**` line said `1.0.0`; it had been stale since 1.1.0.
+  Synced to 1.8.0 and annotated that the schema `description` is the changelog of record. Added one
+  bullet to "Cross-field rules" for the new enum. That was the whole of the staleness; nothing
+  ballooned.
+
+**What the renderer emits** for `hardware.gpu: intel-quicksync`, on the app container only (never a
+sidecar), and it is exactly two things:
+
+1. a hostPath volume `dri` → `/dev/dri`, `type: DirectoryOrCreate`;
+2. a read-write `volumeMount` of `dri` at `/dev/dri` on the app container (deliberately not
+   `readOnly` — VA-API opens the render node read-write).
+
+**Mounting the device is sufficient.** It emits **no** `privileged: true` (Kyverno
+`restrict-host-namespaces` is Enforce on app namespaces and it is not needed), **no** `nodeSelector`
+(placement is a separate concern with its own justification), **no** resource requests (there is no
+app-container resource render path), and — the one that is worth spelling out — **no
+`podSecurityContext.supplementalGroups`**.
+
+**Why no `supplementalGroups`, when `/dev/dri/renderD128` is `root:render 0660`.** An earlier
+revision of this branch merged the host `render` GID into `supplementalGroups`. **It was measured to
+be a no-op and has been removed.** In the running container on prod-2, 2026-09-30:
+
+    PID 1   s6-svscan            Uid: 0     Gid: 0     Groups: 0
+    pid 246 Plex Media Server    Uid: 1000  Gid: 1000  Groups: 100 1000
+
+`/etc/services.d/plex/run` ends in `exec s6-setuidgid plex …`, and the child's group list is exactly
+the `/etc/group` account-database list — root's gid `0`, which PID 1 holds, has **vanished**. That is
+a `setgroups()` that *replaces* rather than appends, so anything Kubernetes puts in
+`supplementalGroups` reaches `s6-svscan` and is discarded before the app ever starts. What actually
+works is the image's own `/etc/cont-init.d/45-plex-hw-transcode-and-connected-tuner`, which `find`s
+every char device under `/dev/dri`, `stat -c '%g'`s its GID, `groupadd`s a matching group inside the
+container when none exists, and `usermod -a -G`s the app user into it — every start, nothing
+hardcoded; `cont-init.d` completes before `services.d`, so the database is correct at exec time.
+
+Shipping an inert `supplementalGroups` would have been worse than shipping nothing: it would read as
+the mechanism to whoever debugs this next. The reasoning lives in a comment **at the emission site**
+and in `renderer-contract.md`, and `render.py --selftest` now asserts the *absence* of a
+`securityContext`, so it cannot be "helpfully" added back without a test failing. A pleasant
+corollary: because the hook re-`stat`s the device every start, host GID drift is **self-healing**,
+and no GID is pinned anywhere in this repo.
+
+**The design decision worth defending: the field is an enum of intent, not a device path.** The SD
+says *what class of host device the workload needs*; the hostPath and the mount exist only in
+`catalog/render.py`. Migrating to the Intel GPU device plugin (`gpu.intel.com/i915` resource
+requests, no hostPath at all) is then a renderer change with zero catalog edits. Both the schema
+`description` and the contract section say this explicitly, because the obvious "simplification" —
+letting an SD name `/dev/dri` — would push a platform implementation detail into the operator-facing
+contract and make that migration a catalog-wide edit.
+
+**`DirectoryOrCreate` is load-bearing.** A node that comes back without the render device (rebuild,
+BIOS change, driver regression, hardware swap) must degrade the workload to software transcoding,
+not fail the mount: `apps-wave-0` is `wait: true` and `apps-wave-1` `dependsOn` it, so a failed
+hostPath mount stalls the entire app tier for the sake of a feature that is not load-bearing.
+`/dev` is tmpfs, so the empty directory the kubelet creates on such a node leaves nothing durable
+behind. Do not "tighten" it to `Directory`. The reasoning is recorded at the emission site.
+
+### Rebase note (2026-10-01, Ruth Okonkwo) — resolution only, no redesign
+
+Rebased onto `main` @ `08deb20` (was cut from `0fbf544`). Re-landed as a single cherry-pick onto
+current `main`; the branch ref was force-updated with `--force-with-lease`, so PR #27 is the same
+PR with a new head.
+
+**One conflict, in `catalog/render.py`, purely textual.** PR #28
+(`fix/oauth2-proxy-cookie-collision-fleet`, merged as `6caa5f2`) inserted its own constants blocks
+(`OAUTH2_ACCESS_TOKEN_LIFETIME_S` / `OAUTH2_COOKIE_REFRESH`, and `OAUTH2_PROXY_METRICS_PORT`) at the
+same insertion point immediately after `KEYCLOAK_ISSUER` that this branch used for
+`DRI_VOLUME_NAME` / `DRI_DEVICE_DIR`. The two blocks are unrelated and do not interact, so **both
+sides were kept in full**, oauth2-proxy's first and the hardware block after it. Nothing was
+dropped and nothing was rewritten: the rebased commit adds the **same 126 `render.py` lines** as the
+pre-rebase commit (verified by comparing the added-line sets) and deletes none. In particular the
+`supplementalGroups`-is-deliberately-not-emitted and `DirectoryOrCreate`-is-load-bearing comments
+survive verbatim.
+
+Re-verified after the resolution: `--selftest` PASS (rc 0), and a full re-render still produces a
+**zero** diff in `apps/` — this PR remains inert, as it must be while no ServiceDefinition opts in.
+
+## What I verified
+
+Measured, in this worktree, on this branch:
+
+- **`python3 catalog/render.py --selftest` → `render.py selftest: PASS`, rc 0.** New cases in the
+  existing `check(...)` style: the positive (volume + mount, and *no* `securityContext`, *no*
+  privileged, *no* nodeSelector), the negative that matters (an SD without `hardware` gains no
+  volume, no volumeMount and no `securityContext` — mirroring the existing `podSecurityContext`
+  negative), and a pass-through case (an SD declaring both `hardware` and its own
+  `podSecurityContext: {fsGroup: 1001, supplementalGroups: [44]}` keeps it **exactly** as written).
+- **The self-tests actually bite.** I mutated the guard (`if hw_gpu == "intel-quicksync":` → `if
+  True:`) in a scratch copy and re-ran: **rc 1**, failing on "an SD without `hardware` must gain no
+  dri volume, no mount and no securityContext". A test that cannot fail is not a test; this one can.
+- **Re-render is a zero diff.**
+  `python3 catalog/render.py --all catalog -o apps --wave-layout --no-image-automation` then
+  `git diff --stat apps/` → **empty**, and `git status --porcelain apps/` → **empty** (so the two
+  not-onboarded by-products, `apps/wave-1/keycloak.yaml` and `apps/wave-0/vm-demo.yaml`, were
+  rewritten byte-identically and needed no revert). The change is inert until a service opts in.
+- **Schema validity and catalog conformance.** `Draft202012Validator.check_schema` passes; all
+  **12** `catalog/*.yaml` validate with **0** errors. On the new field specifically: `gpu:
+  intel-quicksync` accepted; `gpu: /dev/dri` rejected (enum); `hardware: {gpu: …, path: …}` rejected
+  (`additionalProperties: false`); `hardware: {}` rejected (`minProperties: 1`).
+- **Hand-checked the emitted YAML** by rendering a throwaway SD carrying `hardware`, a Longhorn
+  volume and its own `podSecurityContext: {runAsUser: 1000, supplementalGroups: [44]}`. Result:
+  `volumeMounts` = config then `{name: dri, mountPath: /dev/dri}`; `volumes` = the PVC then
+  `{name: dri, hostPath: {path: /dev/dri, type: DirectoryOrCreate}}`; `securityContext` =
+  `{runAsUser: 1000, supplementalGroups: [44]}` — i.e. **unchanged**, nothing appended. That toy SD
+  was never written into `catalog/`.
+- **GATE: `rc 3` — UNDETERMINED, NOT green.** Run in the pinned toolbox from this worktree with
+  both CLAUDE.md §3 workarounds (`DEV_KUBECONFIG=/dev/null`, and an `ENGINE` shim bind-mounting the
+  main checkout's `.git` at its own absolute path). Diff base `origin/main @ 0fbf544d5e2f`; touched
+  requirements `R-DNS-03 R-GITOPS-02 R-IDP-02`. Run **four** times on this branch with an identical
+  result each time, most recently against tree `c132893045b1` — the final commit differs from that
+  run only by this paragraph. Per stage:
+
+  | stage | result |
+  |---|---|
+  | spec-immutability | **SKIPPED** — sibling spec repo not at `/work/../home-network-k3s`. Expected from a worktree; tolerated by §3. |
+  | decisions-log | pass |
+  | lint | pass |
+  | l0 | pass — 19 offline cases, `0 failure(s), 0 skip(s), 0 unwired` |
+  | render-golden | pass — *"all 10 onboarded service golden(s) reproduce from the canonical render"* |
+  | l1 | pass — 3 offline checks, `0 failure(s), 0 skip(s), 0 unwired`; coverage 51/51 requirements scaffolded |
+  | approval-policy | pass (`event='none'`, not a pull_request event — N/A locally; it will evaluate for real on the PR) |
+  | gitea-settings | **SKIPPED** — `SETTINGS_API_TOKEN` unset. Expected; branch protection and runner registration are therefore UNVERIFIED by this run. |
+  | conformance (L2/L3) | **SKIPPED** — `CLUSTER_AVAILABLE!=true`. Phase-0 standing skip. |
+
+  The gate's own verdict: *"Stages that did not run: spec-immutability gitea-settings"*. **No stage
+  failed, and no stage skipped for a reason caused by this change.** rc 3 is reported as rc 3; I am
+  not calling this green.
+- **Blast radius of the schema change:** `hardware` is absent from every catalog entry, so
+  `apply_defaults` never materializes it and `s.get("hardware")` is `None` on all 12 — which is what
+  the zero diff above demonstrates rather than asserts.
+- **Touched no sensitive path.** Changes are confined to `schemas/`, `catalog/`, `decisions/`; per
+  `infra/gitea/approval-policy.json` none of those are sensitive. `conformance/**` and `clusters/**`
+  are untouched. I did not run `--emit-waves` (it writes `clusters/base/`, and the set of files
+  under `apps/` did not change).
+
+## What I did NOT verify
+
+- **Whether containerd's device cgroup permits a non-privileged pod to `open()` a bind-mounted char
+  device. Still unverified, and it is the single largest technical risk in this workstream.**
+  containerd installs a per-container **device cgroup** allow-list, and whether
+  `/dev/dri/renderD128` arriving as a bind-mounted device node is permitted by it on this
+  k3s/containerd version is **unmeasured**. Settling it requires applying a pod, which is an
+  operator action, not an agent one. The container's PID 1 running as uid 0 helps on the DAC side
+  but does **not** bypass the cgroup gate. If it blocks, the mount will succeed, the image's group
+  hook will do its job, and the `open()` will still fail — and the fallback is the Intel GPU device
+  plugin, which *sets* the cgroup entry. Note that that fallback is a **renderer-only change**: no
+  ServiceDefinition moves, because the authoring surface is an intent enum. That is the design
+  justifying itself, and it is the reason to keep the enum even if the hostPath mechanism loses.
+- **That a missing device degrades silently — and nothing will tell you.** This is the accepted cost
+  of `DirectoryOrCreate`. The workload will software-transcode and **no signal reports it**: media
+  apps in this catalog declare no `metricsEndpoint`, so no ServiceMonitor exists, and there is no
+  metric to alert on. I deliberately did **not** propose an alert, because an alert that cannot fire
+  is worse than none — see the 2026-09-13 correction in `renderer-contract.md` §`observability`,
+  where twelve monitors discovered zero targets. The only proof of hardware transcode is an operator
+  seeing it in the application's own dashboard.
+- **That "mounting the device is sufficient" generalizes beyond this one image.** It does not, and
+  I have not verified that it does. The claim was measured for an image that resolves the device
+  group itself in `cont-init.d`. A future consumer of `hardware.gpu` that runs non-root with no
+  equivalent hook would still need the group solved — and the fix for that case is a **renderer**
+  change (per-image handling, or the device plugin), never a new SD field. I scoped the claim in
+  `renderer-contract.md` rather than letting it read as general; the first such consumer must
+  re-measure.
+- **The feature now depends on a hook inside an image that floats.** The `cont-init.d` behaviour
+  above was read from digest `sha256:e0ab27395614…` (`plexmediaserver 1.43.4.10903`), but the
+  deployment runs `plexinc/pms-docker:latest` with `imagePullPolicy: Always`. **A silent pull can
+  change the behaviour this feature depends on, with no commit anywhere in this repo to show for
+  it** — and the failure would be the silent software-transcode described above. I have not fixed
+  that here (it is the Plex wiring's tag/digest decision, not the capability's), but it should not
+  be forgotten.
+- **GID drift is NOT on this list any more, deliberately.** An earlier revision pinned
+  `INTEL_RENDER_GID = 992` and listed "a node rebuild could move it" as an open risk. The hook
+  re-`stat`s the device on every start, so drift is self-healing and no GID is pinned anywhere. The
+  risk is retired, not overlooked.
+- **Anything on the live cluster.** I ran no `kubectl` and I opened no shell in any container. The
+  node facts (device present, `root:render` 0660, `i915` bound, no device plugin, no
+  `gpu.intel.com/*` advertised), the in-container measurements quoted above, and the Kyverno
+  posture (`restrict-host-namespaces` Enforce; no hostPath or `supplementalGroups`
+  restriction; no PSA labels on app namespaces) are **prior recon handed to me, not re-measured by
+  me**. Per CLAUDE.md §2b, treat them as second-hand until re-queried.
+- **That CI exercises the self-tests. It does not.** `git grep -- --selftest ci/` returns nothing —
+  no gate stage invokes the renderer's self-test. They are the executable spec and house style, and
+  I ran them by hand, but a future regression in this code path **will not be caught by the gate**.
+  Wiring `--selftest` into a gate stage is a real gap and is not part of this PR (`ci/` is a
+  sensitive path and would drag in a human reviewer for a change unrelated to this one).
+- **Runtime behaviour of any kind.** The gate proves the L0/L1 offline subset only;
+  `vars.CLUSTER_AVAILABLE` is unset, so L2/L3 conformance does not run in CI.
@@ -0,0 +1,256 @@
+---
+date: 2026-10-01
+branch: fix/admin-approval-override
+title: Administrators can merge their own green PR — required_approvals 1 → 0, plus an attributed /approve-override for sensitive paths
+---
+
+## What changed
+
+**The problem, measured first.** `required_approvals: 1` on `main` was a lock-out, not a
+protection. Gitea refuses to let a pull request's author approve their own PR, and `jakman85` is
+the **only** collaborator (permission `owner`). So no account in existence could approve a PR the
+operator opened. Measured on the live API, 2026-10-01:
+
+| | |
+|---|---|
+| PR #35 (`bugfix/arr-to-plex`, author `jakman85`) | `gate / gate (pull_request)` = **success, 1m41s** |
+| its reviews | `[]` |
+| live `required_approvals` | `1` |
+| live `block_admin_merge_override` | `true` |
+
+Green gate, zero reviews, no path to a merge — and the same was true of any PR that would have
+fixed it.
+
+It was also never the design this repo wrote down. `ci/approval-policy.sh`'s own header and
+`ci/gate.sh:194` both already described the model as *"`required_approvals` drops to 0 and the
+GATE decides"*, and `infra/gitea/require-gate.sh`'s closing note named this exact failure and
+prescribed this exact remedy. `decisions/2026-09-30-flux-deploy-notification-inhibited.md` had
+flagged the 1-vs-0 disagreement as unresolved and not that session's to settle. This settles it
+at 0.
+
+**Two layers were blocking, and both are addressed.**
+
+*Layer 1 — Gitea branch protection.* `required_approvals: 1 → 0` in
+`infra/gitea/expected-branch-protection.json`, with the rationale in the file, plus
+`infra/gitea/allow-admin-self-merge.sh` to apply it to the live rule (it reads the target **from
+the declaration**, so the script cannot drift from what the gate checks). `require-gate.sh` was
+changed to 0 as well — left at 1 it would silently re-impose the lock-out on its next run.
+
+The narrow choice matters: the *other* way to unblock an owner is
+`block_admin_merge_override: false`, and that would have re-opened precisely what
+`infra/gitea/block-admin-override.sh` was written to close (PRs #5–#9 merged with zero approvals
+while the settings read as enforced). This change drops the approval **count** only.
+`block_admin_merge_override` and `enable_status_check` both stay `true`, and the script asserts
+both afterwards and fails loudly if the PATCH disturbed either. **A red gate remains unmergeable
+by everyone, administrators included.**
+
+*Layer 2 — `ci/approval-policy.sh`.* Sensitive paths (`ci/**`, `infra/gitea/**`, `CLAUDE.md`,
+kyverno, rbac, sops, Flux wiring) still require a non-author `APPROVED` review, which the sole
+administrator still cannot obtain. An administrator may now waive it by commenting on the PR:
+
+    /approve-override <reason>
+
+- the directive must **start a line** (a quoted or mid-sentence mention does not count, or
+  repeating someone else's override would re-arm it);
+- a **reason is mandatory** — a bare directive records that review was waived but not why;
+- the commenter's permission is **re-checked live** against
+  `/collaborators/{login}/permission` when the gate runs, so an old comment from a demoted
+  account carries nothing;
+- it is **invalidated by a new commit** (the comment must post-date the head commit), mirroring
+  Gitea's own `dismiss_stale_approvals`;
+- the overrider, timestamp and reason are printed in the gate log and the job summary, and a
+  self-waiver by the PR's own author is called out as such.
+
+Selection of the candidate comment is `ci/approval-override.py` — pure, offline, no authority of
+its own — so it could be unit-tested; authorisation stays in the fail-closed stage.
+`ci/tests/approval-override.test.sh` adds **23 assertions**, all passing, each constructing a
+case the parser must refuse. Every new failure mode blocks: unreadable head-commit timestamp,
+unreadable permission, parser error.
+
+**The cost, stated rather than buried.** For an administrator, *"a change to the gate cannot be
+gated by the gate"* is now **advisory, not structural**. That is unavoidable with one human — the
+alternative is a repository its owner cannot change — and what is bought back is an attributed,
+reasoned, commit-scoped record of each waiver instead of a silent bypass.
+
+**The uncomfortable part, accepted with eyes open (operator decision).** Authorisation is by live
+permission **alone**, with no allowlist of human accounts. The **agent account `gitea-admin` also
+reports permission `owner`** on this repo (measured), so an agent *can* clear the sensitive-path
+review on its own PR — and agents author most PRs here, so that is the common case, not an edge
+case. I raised this before implementing; the operator chose permission-only. It is documented in
+three places rather than left to be discovered: `approval-policy.json`
+(`_known_consequence_agent_accounts`), the stage's header, and `CLAUDE.md`, which now tells
+agents plainly **not** to use it on their own work. The narrowing path needs no code change —
+add an `override_admins` list to `approval-policy.json` and the stage already requires membership
+in addition to the permission check.
+
+Also updated: `CLAUDE.md` §1 and the §2 enforcement table (the "approval required" row is now
+explicitly ⬜ with the reason, rather than a stale ✅), and `infra/gitea/README.md` gains §6 and
+marks its historical step 4 as reversed — do not re-apply.
+
+**Second pass, 2026-10-02 — the override did not work the first time, and why.** The operator ran
+the live PATCH (`required_approvals: 0` confirmed live, with `block_admin_merge_override` and
+`enable_status_check` still `true`), merged PR #35, and posted `/approve-override Need to ensure
+approval override work as expected` on PR #38. The gate re-ran (run 203) and **still blocked.**
+
+The comment was not at fault: posted by `jakman85`, directive at the start of the line, reason
+present, and 29 minutes **newer** than head commit `616a497`, so it cleared the staleness rule.
+What failed was that **CI could not read it.** Run 203's log:
+
+    approval-override: COMMENTS_JSON is not a list
+
+`SETTINGS_API_TOKEN` is minted `["read:repository", "read:admin"]` by
+`infra/gitea/set-ci-token.sh`, and Gitea puts `/issues/{n}/comments` behind the **issue** scope. So
+Gitea returned an error *object*, `jq -e .` accepted it as valid JSON, and the stage fell through
+to "no approving review" — **a permission failure wearing the costume of a missing override.**
+`/pulls/{n}/files` and `/pulls/{n}/reviews` are under `read:repository`, which is exactly why every
+other part of the same stage worked and this one did not.
+
+Three fixes, and the first is the one that mattered most:
+
+1. **The diagnosis is no longer misleading.** `COMMENTS_JSON` is out of the must-parse-or-block
+   loop and is explicitly checked for being an *array*; when it is not, the stage prints Gitea's
+   own `.message`, names the missing scope, and says where to post the override instead. An
+   unreadable source can only cause a false BLOCK, never a false pass, so degrading it to a loud
+   warning is safe — and blocking on it was never the honest behaviour.
+2. **The override is accepted from a PR REVIEW body as well as a comment**, so the hatch is not
+   hostage to one API scope. A review of type "Comment" works on the token CI has *today*. The
+   gate names which source it used. `PENDING` (unsubmitted, author-only draft) reviews are dropped.
+3. **`set-ci-token.sh` now requests `read:issue`**, which makes the ordinary comment box work —
+   the thing anybody reaches for first. It takes effect only on a **re-run**: an existing token's
+   scopes cannot be widened in place.
+
+`ci/tests/approval-override.test.sh` grows from 23 to **38 assertions**, covering the reviews
+source, `PENDING` reviews, an error *object* from the comments API, "an unreadable comments source
+must not suppress a valid review override", latest-wins across both sources, and that the
+diagnostic actually quotes Gitea's message.
+
+**One adjacent fix, deliberately in scope: `ci/tests/gate-decision.test.sh` was already broken.**
+It hand-maintains a list of stub stage scripts that "must name EVERY stage `gate.sh` runs", and it
+did not name `approval-policy` — added to `gate.sh` by PR #17 and never added here. `run_stage`
+therefore got exit 127 from a missing file and recorded a FAILED stage in every case, so **15 of
+50 assertions failed**, including every one expecting a PASS/UNDETERMINED headline. Measured on
+pristine `origin/main` (`0c7e8dd`) before touching it, so the attribution is not a guess.
+
+I fixed it because `make -C ci test` is the *only* runner for the suite this change adds — leaving
+it red would have made the new tests unverifiable in practice. The file's own comment already
+narrated this exact failure for `gitea-settings` and did not prevent the repeat, which says the
+comment was not the fix. So the stub list is now **derived from `gate.sh`'s own `run_stage` lines**
+and asserts it parsed at least 8 of them: a stage added to the gate can no longer be silently
+missing from its test. `make -C ci test` goes from **FAILED to all suites passed**.
+
+## What I verified
+
+- **The lock-out is real, and PR #35 is the proof.** Queried live: head sha
+  `8a094c46`, statuses `gate / gate (pull_request) = success :: Successful in 1m41s`,
+  `/pulls/35/reviews` → `[]`, `/branch_protections/main` → `required_approvals: 1`. I read the
+  **rule object**, not `/branches/main`, per §2b.
+- **`jakman85` is the only collaborator**: `/collaborators` → one entry. Permission `owner`.
+- **`gitea-admin` also reports `owner`** — this is why the agent-account consequence above is a
+  statement of fact and not a hypothetical.
+- **PR #35 touches no sensitive paths** (`apps/wave-0/plex.yaml`, `apps/wave-1/{radarr,sonarr}.yaml`,
+  `catalog/{radarr,sonarr}.yaml`, its decisions entry) — so `ci/approval-policy.sh` was never its
+  blocker. Layer 1 alone unblocks it.
+- **API shapes verified before depending on them**, rather than assumed: `/git/commits/{sha}`
+  returns `.commit.committer.date` as RFC3339 **with an offset** (`2026-10-01T23:09:08-05:00`),
+  not a trailing `Z` — a parser handling only `Z` would have silently dropped every real
+  override, so there is a test for the offset form. PR conversation comments are under
+  `/issues/{n}/comments`; `/pulls/{n}/comments` returns diff comments only and would never have
+  seen an override.
+- **`ci/tests/approval-override.test.sh`: 23 passed, 0 failed.** Covers the happy path, bare
+  directive, mid-sentence and `>`-quoted directives, stale and equal-timestamp overrides,
+  unparseable timestamps, missing login, invalid JSON, non-list JSON, latest-wins, offset
+  timestamps, and a configurable directive.
+- `bash -n` clean on `ci/approval-policy.sh`, `infra/gitea/allow-admin-self-merge.sh`,
+  `infra/gitea/require-gate.sh`; `jq -e .` clean on both JSON declarations.
+- **`gate-decision.test.sh`'s 15 failures are pre-existing, measured not assumed.** Extracted
+  `origin/main`'s `ci/` at `0c7e8dd` into a throwaway directory and ran the suite from there:
+  **FAIL, 15 failed / 35 passed**. After deriving the stub list from `gate.sh`: **PASS, 50
+  assertions**.
+- **`make -C ci test` in the pinned toolbox: `ci/tests: all suites passed`, rc 0** — approval-override
+  **38**, gate-decision 50, touched-requirements-base 31, test-attribution 15, touched-requirements 26,
+  decisions-log pass. It was rc 2 (`ci/tests: FAILED`) before the stub-list fix.
+- **The live flip worked, and the drift detector proved it.** Run 202 (before): `gitea-settings:
+  FAIL`, `required_approvals: DECLARED 0 but LIVE is 1`. Run 203 (after the operator ran the
+  PATCH): `gitea-settings: pass`, with `block_admin_merge_override = true` among the keys it
+  checked. So the narrow change held — the approval count moved and the gate's teeth did not.
+- **The first override attempt failed for a measured reason, not a guessed one.** Run 203 log:
+  `approval-override: COMMENTS_JSON is not a list`. Token scopes read from `set-ci-token.sh`:
+  `[read:repository, read:admin]`. The same run successfully read `/pulls/38` and
+  `/pulls/38/files` (author and sensitive paths both resolved) and got past the reviews
+  validation, which is how I know `/pulls/{n}/reviews` IS readable with that token — from
+  behaviour, not from documentation.
+- **The comment itself was valid.** `created_at` 2026-10-02T05:31:24Z vs head commit `616a497`
+  committer date 2026-10-02T00:02:20-05:00 → the comment is 29m04s newer, so it passed the
+  staleness rule. Ruling the comment out is what left the token scope as the only candidate.
+- **`?limit=1000` is not the culprit.** Queried `/issues/38/comments` with `limit=1000`, `limit=50`
+  and no limit using my own credentials: `type=array` in all three. So the difference is the
+  credential, not the query.
+- **The root cause is confirmed by Gitea, in its own words.** Run 206 (the fix in place):
+  `[warn] cannot read PR COMMENTS from Gitea — token does not have at least one of required
+  scope(s), required=[read:issue], token scope=read:admin,read:repository.` followed by the
+  instruction to post the override as a review. That is the diagnosis the first attempt should
+  have produced, and `gitea-settings: pass` in the same run. The stage still ends `FAIL`, correctly
+  — there is no valid override on the PR yet.
+- **Toolbox merge gate** (`./infra/devcontainer/dev.sh run bash -c './ci/gate.sh'`), run on the
+  rebased branch: **rc 3, `MERGE GATE: UNDETERMINED`**. Green stages: spec-immutability,
+  decisions-log, lint, l0 (23 offline cases, 0 failures), render-golden including the new
+  `catalog/render.py --selftest`, l1 (3 cases), approval-policy (N/A offline). Skipped:
+  `gitea-settings` (no `SETTINGS_API_TOKEN` on this host) and `conformance` (no cluster).
+  **This is NOT a green run and I am not reporting it as one** — rc 3 is UNDETERMINED per the
+  CLAUDE.md §3 contract.
+
+## What I did NOT verify
+
+- **The override has still never been ACCEPTED by a real run.** This is the second attempt at
+  this bullet, which is itself the point: the first pass called the hatch unverified, and the
+  first real test found a defect that no amount of local testing would have (an API scope CI has
+  and I did not). What is now measured is that the parse → *reject* path works end to end and
+  diagnoses itself. The **accepting** branch — directive found → live permission lookup → exit 0 —
+  has never executed. Posting the override as a **review** on PR #38 is its first real test, and
+  I would not be surprised by a third round.
+- **`read:issue` is the right scope name — Gitea said so itself, so this is no longer a guess.**
+  Run 206 printed Gitea's own message: `token does not have at least one of required scope(s),
+  required=[read:issue], token scope=read:admin,read:repository`. That both confirms the root
+  cause and names the string, which matters because this repo has been burned by a token *name*
+  that could not exist (`GITEA_SETTINGS_TOKEN`). What is **still** unverified is that the token
+  *mint* accepts it — `POST /users/{u}/tokens` with that scope has not been run. The review route
+  does not depend on it.
+- **I did not verify that Gitea lets a PR's own author submit a `COMMENT`-type review.** Gitea
+  blocks author self-*approval*; "Comment" is expected to be allowed, and that expectation is now
+  load-bearing for the recommended route. It is one click to find out, and I deliberately did not
+  test it by posting an override myself: doing so as `gitea-admin` is precisely the agent
+  self-override that `CLAUDE.md` now forbids.
+- **The whole re-run mechanism is manual.** `gate.yml` triggers on `pull_request` and `push: main`
+  only — there is no `issue_comment` trigger, so posting an override does **not** re-run the gate.
+  Re-run the job from the Actions UI, or close and reopen the PR (which fires `pull_request` and
+  leaves the head SHA alone). **Do not push an empty commit to re-trigger it:** that moves the head
+  commit and makes the override stale by the rule above, so it would have to be posted again. A
+  comment trigger was considered and not built — it would need to check out PR head code while the
+  job holds `SETTINGS_API_TOKEN`, which is the classic pwn-request shape and not worth it here.
+- **I have not seen the gate still REFUSE a merge** now that approvals are 0. The claim that
+  `block_admin_merge_override: true` keeps a red gate unmergeable rests on Gitea's documented
+  behaviour and on the 2026-09-29 session, **not** on a provoked refusal in this one. That probe
+  is written out at the foot of `allow-admin-self-merge.sh` and should be done, exactly as the
+  push rejection was provoked rather than assumed.
+- **Clock skew in the staleness check is unhandled.** Staleness compares the comment's
+  server-side `created_at` against the head commit's *committer* date, which comes from the
+  committer's machine. A commit dated in the future would reject legitimate overrides until real
+  time catches up (fail-closed, so loud, but annoying). Gitea's own `dismiss_stale_approvals`
+  compares SHAs instead; a comment cannot. Not hit here — PR #35's commit date is ~21 min behind
+  server time — and not fixed.
+- **`ci/tests/approval-override.test.sh` does not run in the merge gate.** `ci/gate.sh` runs no
+  `ci/tests/*`; they run only via `make -C ci test`. So a regression in the parser would not be
+  caught by CI. PR #33 (merged while this branch was being written, and rebased onto here) fixed
+  the *same class* of gap for `catalog/render.py --selftest` by invoking it from
+  `ci/render-golden.sh` — it did **not** wire `ci/tests/*` in, which I confirmed after rebasing
+  rather than assuming. Wiring them is a one-line stage and I did not widen scope to it.
+- **Authorisation itself is untested code.** `ci/approval-override.py` is unit-tested; the live
+  permission lookup, the `override_admins` allowlist branch, and every new BLOCK path in
+  `ci/approval-policy.sh` need Gitea and have never been exercised. The allowlist branch in
+  particular is implemented but **unused** — `override_admins` is absent by design.
+- **The other open PRs will go red** once the live value changes, until they are rebased onto the
+  merged declaration: as of writing those are **#29** and **#37** (both `gitea-admin`), plus
+  **#35**. I did not rebase them. Merging them first avoids it entirely; the script's header gives
+  the order. PRs #33, #34 and #36 merged while this branch was being written — this branch is
+  rebased onto `0c7e8dd`, so the list was re-measured rather than carried forward.
+- No cluster state was read or changed; nothing here touches Flux, Kyverno or any workload.
@@ -0,0 +1,120 @@
+---
+date: 2026-10-01
+branch: bugfix/arr-to-plex
+title: Radarr and Sonarr may now reach Plex on 32400 — the network path only, not the Plex connection itself
+---
+
+## What changed
+
+Radarr and Sonarr had no route to Plex. The media library lives on NFS
+(`/mnt/Storage/plex`, the ZFS dataset `Storage/plex` on `${BULK_NFS_SERVER}`), and NFS
+does not deliver filesystem CRUD notifications, so Plex never learns that an arr has
+written a new file. Without a push from the arrs, the library only reflects new media
+when somebody scans it by hand. That is the bug this branch addresses.
+
+The fix is one `dependencies:` entry per app, added to the catalog — the source of truth:
+
+    - to: plex
+      port: 32400
+      protocol: tcp
+
+in both `catalog/radarr.yaml` and `catalog/sonarr.yaml` (these two *are* the arr suite
+here; `jackett`, `qbittorrent` and `seerr` are different roles and are untouched).
+
+The renderer expands each entry into a **pair** of Cilium policies, which is why a
+three-line catalog change produces 88 lines under `apps/`:
+
+| Side | Policy | File |
+|---|---|---|
+| egress | `radarr-allow-plex-32400` | `apps/wave-1/radarr.yaml` |
+| egress | `sonarr-allow-plex-32400` | `apps/wave-1/sonarr.yaml` |
+| reciprocal ingress | `plex-allow-radarr-32400` | `apps/wave-0/plex.yaml` |
+| reciprocal ingress | `plex-allow-sonarr-32400` | `apps/wave-0/plex.yaml` |
+
+Three commits, and the middle two are worth recording because both were self-inflicted
+gate failures rather than design changes:
+
+- `753c48b` — the catalog edit alone, with `apps/` left unrendered.
+- `9048c2f` — re-rendered `apps/` from the catalog, which `render==golden` requires.
+- `58f02fe` — removed 20 trailing spaces on `catalog/radarr.yaml:103`, left behind as
+  comment-alignment padding when the comment moved down to the `port:` line. Sonarr's
+  equivalent line kept its comment in place, which is why only radarr tripped.
+
+## What I verified
+
+**The lint failure, root-caused.** Gate run 193 on `9048c2f` failed with
+`[fail] lint: FAILED -> blocks merge`, and lint was the only failing stage. yamllint
+emitted 419 findings, but **418 were `line-length`**, which `ci/yamllint.yaml` pins at
+`level: warning` and which therefore cannot change the exit code. The single error-level
+finding in the entire repository was:
+
+    catalog/radarr.yaml
+      103:13    error    trailing spaces  (trailing-spaces)
+
+`trailing-spaces` inherits the default severity (error), so that one line returned rc 1.
+After the fix, `git grep -nIE ' +$' -- '*.yaml' '*.yml'` matches nothing repo-wide.
+
+**Gate run 195 on `58f02fe`** (read from the job log, not from the status tick):
+
+    [ ok ] approval-policy: green     [ ok ] l0: green
+    [ ok ] gitea-settings: green      [ ok ] render-golden: green
+    [ ok ] decisions-log: green       [ ok ] l1: green
+    [ ok ] lint: green
+    [fail] MERGE GATE: UNDETERMINED — greenness was NOT ESTABLISHED.
+    [fail]   Stages that did not run: spec-immutability
+
+The Gitea check reports `success` / "Successful in 1m41s", but the gate's own verdict is
+**UNDETERMINED (rc 3)**, not PASS. That tick is `gate.yml` tolerating exit 3 for the one
+stage it documents as unverifiable on this runner (the spec repo is not checked out
+there). Nothing else produced a 3.
+
+**`render==golden` green** on `58f02fe`, so `apps/` is a faithful canonical render of
+`catalog/` — the four generated policies are not hand-written drift.
+
+**DNS egress was already open** and is not a new gap: `apps/wave-1/radarr.yaml` carries
+an existing allow for port 53 UDP+TCP, so resolving the Plex service name does not need
+anything added here.
+
+**Live branch protection on `main`**, measured against
+`/api/v1/repos/Home-Network/home-network-k3s-impl/branches/main` rather than assumed:
+`protected: true`, `user_can_push: false`, `enable_status_check: true`,
+`status_check_contexts: ['gate / gate (pull_request)']`, `required_approvals: 1`.
+
+## What I did NOT verify
+
+**This change does not accomplish what the PR set out to do.** It opens the network path
+and nothing else. Triggering a library update also requires each arr's
+*Settings → Connect → Plex Media Server* entry plus a Plex token, and **none of that
+exists anywhere in `catalog/` or `apps/`** — I grepped both for `plex`, `token` and
+`connect` and found only NFS paths, a uid and a timezone. Radarr's and Sonarr's config
+lives on a 5Gi Longhorn PVC mounted at `/config`, so that connection is UI-only state
+that exists solely on the PVC and never in git, in the same way Crafty's local login
+does. **Until somebody adds it in each app's UI, Plex still will not see library
+updates, and this branch will look like it did nothing.** That follow-up is not tracked
+by any file in this PR.
+
+**L2/L3 conformance never ran.** `CLUSTER_AVAILABLE` is unset on the runner, so the
+eight requirements this diff touches — `R-NET-01`, `R-NET-05`, `R-NET-06`, `R-NET-07`,
+`R-IDP-02`, `R-IDP-06`, `R-GITOPS-02`, `R-PLAT-02` — are **UNVERIFIED**. That gap is
+unusually load-bearing here: the entire substance of this change is four
+CiliumNetworkPolicies, and no test has confirmed that traffic from the `radarr` or
+`sonarr` namespaces actually reaches `plex:32400`. The green gate proves the offline
+render subset only.
+
+**`spec-immutability` did not run** (`R-TEST-06` not verified), for the documented
+reason — the sibling spec repo is absent from the runner.
+
+**Nothing was checked on the cluster.** I ran no `kubectl`, so I did not confirm that
+the `plex` Service exposes 32400 as the arrs would address it, that the pod labels the
+policies select (`app.kubernetes.io/name`, `k8s:io.kubernetes.pod.namespace`) match the
+running pods, or that Cilium accepted the four new objects. Flux has not applied any of
+this: the branch is unmerged, so `.status.lastAppliedRevision` does not contain it.
+
+**I did not verify Plex's own side** — whether its library-update endpoint on 32400 is
+the right target for the arrs' notification, or whether Plex requires
+"Update my library automatically" enabled to act on one.
+
+**I am not the approver and did not approve this PR.** `required_approvals: 1` is live,
+and the only credential on the authoring workstation belongs to the PR's author
+(`jakman85`), whose own review Gitea does not count. Per CLAUDE.md §1 the merge is a
+human action regardless.
@@ -0,0 +1,198 @@
+---
+date: 2026-10-01
+branch: feat/intel-gpu-device-plugin
+title: Intel GPU device plugin, in its own Kustomization — the resource exists; nothing has run yet
+---
+
+## What changed
+
+Deployed the **Intel Device Plugin for Kubernetes (GPU)**, standalone DaemonSet, pinned at
+**v0.37.0**, so the prod nodes advertise `gpu.intel.com/i915`. This is the measured fix for the
+Plex hardware-transcoding blocker: the `/dev/dri` hostPath merged on `main` @ `cdf3e14` makes the
+device node *visible* and not *openable* (`EPERM` from the device cgroup at both uid 0 and uid
+1000; the container's runc `config.json` carries one deny-all rule and `linux.devices: None`). A
+hostPath is a filesystem operation and adds no cgroup allow rule. A device plugin's `Allocate`
+response is what makes kubelet populate `linux.devices` *and* the matching cgroup allow rule, and
+it is the only non-privileged way to get there — `privileged: true` is blocked by Kyverno
+`restrict-host-namespaces` under Enforce in app namespaces, and `plex` is labelled
+`app-namespace=true`.
+
+New files:
+
+- `platform/intel-gpu-plugin/intel-gpu-plugin.yaml` — the DaemonSet, **vendored** as
+  `kubectl kustomize` output of upstream `deployments/gpu_plugin/base` at `v0.37.0`, under a
+  provenance header carrying the exact re-vendor command and both sha256s (generated body
+  `80ebd070…`, raw upstream source `c8773918…`). Generated rather than byte-copied because
+  upstream mixes YAML sequence-indentation styles within the one document and `ci/yamllint.yaml`
+  sets `indent-sequences: consistent` — a verbatim copy is a hard lint error that blocks the gate.
+  That is not a guess: the first gate run of this branch returned **rc 1 BLOCKED** on exactly that
+  error, and regenerating through kustomize is the fix (and is also the re-vendor idiom
+  `platform/external-snapshotter` documents). Vendored at all, rather than referenced as a remote
+  kustomize base, for the other reason `external-snapshotter` records: a remote base is
+  `git fetch`ed on every reconcile and those fetches have flaked a whole tier.
+- `platform/intel-gpu-plugin/namespace.yaml` — a PLATFORM namespace, explicitly
+  `homelab.keithjackson.dev/app-namespace: "false"` (following `vpn-egress`/`kyverno`).
+- `platform/intel-gpu-plugin/kustomization.yaml` — namespace scoping plus the image pin.
+- `platform/intel-gpu-plugin/README.md` — the measurement, the placement argument, the
+  non-choices, and the post-merge verification sequence.
+- `clusters/base/intel-gpu-plugin.yaml` — **its own Flux Kustomization**: `dependsOn: [cilium]`,
+  `wait: true`, `interval: 10m`, `timeout: 10m`, `prune: true`.
+- One wired line (plus its rationale comment) in `clusters/base/kustomization.yaml`.
+
+**Placement: its own Kustomization, not a line in the shared `platform` tier.** The shared tier was
+cheaper and needed no human reviewer on `approval-policy.json`'s terms, but `platform` is
+`wait: true` over eleven-plus components and the repo already argued this case when Cilium was
+split out: its Ready condition cannot distinguish "the plugin is wedged" from "longhorn is slow",
+the only emergency valve is freezing the entire tier, and the tier's 10m/5m interval/timeout are
+called out there as *wrong for a 3-node DaemonSet rollout* — which is exactly what this is. The
+consequence is worse than average here, because `platform` gates `platform-config` gates
+`apps-wave-0` (`wait: true`) gates `apps-wave-1`: a plugin that never went healthy inside that tier
+would stall every app, so a component added to fix a **silent no-op** would have gained the ability
+to cause a **real outage**. The approval cost of touching `clusters/**` is nil, because
+`required_approvals: 1` with `block_admin_merge_override: true` applies to every PR here regardless
+of path and authors cannot self-approve.
+
+`dependsOn: [cilium]` is the minimum *and* the earliest: the plugin never uses the network and
+never talks to the API server (it talks to the local kubelet over a unix socket), so a working CNI
+is its only real prerequisite, and `cilium` already depends on `flux-system`, which applies the
+`cluster-vars` ConfigMap this Kustomization substitutes from. Depending on `cilium` rather than
+`platform` also means it reconciles *in parallel with* the platform tier rather than behind all
+eleven of its components, which shrinks the cold-start window in which a node has not yet
+advertised the resource.
+
+`wait: true` is a deliberate departure from the `headlamp`/`crafty` leaf template. Those use
+`wait: false` because a leaf UI can be legitimately not-Ready for a long time during a cold start,
+making its health a useless signal. This DaemonSet reads sysfs and opens a socket, so not-Ready
+genuinely means broken — and since **nothing in the graph `dependsOn` it** (verified against the
+rendered graph), `wait: true` gates nothing while turning
+`flux get kustomization intel-gpu-plugin` into a one-line answer to the question that must be
+answered before the consumer PR may merge.
+
+Deliberate non-choices, each argued in the README: standalone DaemonSet rather than the operator
+(no CRD, no controller-manager, no cert-manager webhook, no second namespace to keep healthy);
+`--shared-dev-num` left at the upstream default of 1 (a higher count is a lie the scheduler
+believes — there is no GPU-side partitioning behind it, so it converts a loud "Unschedulable" into
+a quiet "transcodes are slow"); `--health-management` left OFF, because it is the one flag that
+*would* open a device and would reintroduce the exact `EPERM` this fixes; no NFD overlay (all three
+nodes are identical); no `supplementalGroups` anywhere, because the `pms-docker` cont-init hook
+already adds the `plex` user to each device's group and that works identically on a
+plugin-injected device.
+
+**Scope held deliberately:** `catalog/render.py`, `catalog/plex.yaml` and `apps/wave-0/plex.yaml`
+are untouched. Switching Plex from the hostPath to a `gpu.intel.com/i915: 1` request is a strictly
+later PR, and **it must not land until this plugin is merged AND confirmed advertising the resource
+on all three nodes** — a request for a resource no node advertises makes the Plex pod
+Unschedulable, and `apps-wave-0` is `wait: true` with `apps-wave-1` depending on it, so that would
+stall the whole app tier and turn today's silent no-op into a real outage.
+
+## What I verified
+
+Mechanism, read from the v0.37.0 source rather than assumed:
+
+- **The plugin only lists devices; it never opens one.** The scan path uses `os.ReadDir`,
+  `os.ReadFile`, `os.Stat` and `os.Readlink` over `/sys/class/drm` and `/dev/dri`. `grep` for
+  `os.Open|OpenFile|ioctl|syscall|unix\.` across `cmd/gpu_plugin/gpu_plugin.go` (908 lines) returns
+  **nothing**. Listing and `stat()` are what demonstrably work today, so the plugin needs no cgroup
+  grant and no extra privilege — it does not need more than the upstream manifest gives it.
+- **The fix mechanism is real:** `createDeviceSpecsFromDrmFiles` emits
+  `pluginapi.DeviceSpec{HostPath, ContainerPath, Permissions: "rw"}` per DRM node, which is the
+  thing kubelet turns into `linux.devices` plus a cgroup allow rule.
+- **The resource name is `gpu.intel.com/i915` by derivation, not by hope:** `scan()` calls
+  `devTree.AddDevice(devProps.driver(), ...)`, so the suffix is the kernel driver name, and these
+  nodes have `i915` bound.
+- `--shared-dev-num` default is `1` (`flag.IntVar(..., 1, ...)`); `--health-management` default is
+  `false` and `setupLevelZeroService` returns early unless it is set.
+- Missing `/dev/dri/by-path` is non-fatal (`newDevicePlugin` warns, sets `bypathFound = false`).
+
+Image pin, verified against the registry:
+
+- `docker.io/intel/intel-gpu-plugin:0.37.0` exists; manifest digest
+  `sha256:7001200071db10587fc31184eedda4f5cf291efba737adf85a8c80539f7fa5e1` (single amd64
+  manifest, not an index), fetched from `registry-1.docker.io` on 2026-10-01. `v0.37.0` is the
+  latest upstream release (2026-09-16) and is not a prerelease.
+- The Docker Hub tag list carries `0.35.0 / 0.36.0 / 0.37.0` — **image tags have no `v` prefix**
+  while GitHub release tags do.
+
+Policy scope, read from the policies rather than assumed:
+
+- All four app-scoped Kyverno policies (`restrict-host-namespaces`, `require-egress-class-label`,
+  `require-auth-mode-label`, `require-default-deny-baseline`) are gated on
+  `namespaceSelector`/`selector` `matchLabels: homelab.keithjackson.dev/app-namespace: "true"`.
+  The new namespace is labelled `"false"`, so none of them match it.
+- Measured on prod: `plex` **is** `app-namespace=true`, confirming `privileged: true` is genuinely
+  unavailable there; `kyverno` and `vpn-egress` are the two namespaces carrying the label as
+  `"false"`; the other platform namespaces leave it unset. No `intel*` namespace exists, so no
+  collision.
+
+Cluster before-state (read-only, `kubectl` in the pinned toolbox with the prod kubeconfig
+bind-mounted read-only — never copied into the worktree):
+
+- Three nodes `k3s-prod-{1,2,3}`, each `v1.36.4+k3s1`, Debian 13 trixie, `containerd://2.3.4-k3s1.36`.
+- Every extended resource currently advertised is `devices.kubevirt.io/{kvm,sev,tun,vhost-net}` on
+  all three nodes. **No node advertises `gpu.intel.com/*`** — queried explicitly, returned
+  "NONE". That is the before-state this change is meant to move.
+- Those `devices.kubevirt.io/*` resources are themselves advertised by a device plugin, which is
+  measured evidence that the device-plugin API and the default kubelet socket path
+  (`/var/lib/kubelet/device-plugins`) work on these nodes.
+
+Offline rendering (the main proof available to a PR):
+
+- `kubectl kustomize platform/intel-gpu-plugin` renders two objects: `Namespace/intel-gpu-plugin`
+  carrying `app-namespace=false`, and `DaemonSet/intel-gpu-plugin` in namespace
+  `intel-gpu-plugin`, with the image resolved to `docker.io/intel/intel-gpu-plugin:0.37.0`.
+- `kubectl kustomize clusters/base` renders, and the new Flux Kustomization appears as
+  `intel-gpu-plugin  ./platform/intel-gpu-plugin  wait=true  dependsOn=cilium`. Both overlays
+  (`clusters/prod`, `clusters/dev`) also render and carry it.
+- Across the whole rendered graph, **no Kustomization lists `intel-gpu-plugin` in `dependsOn`** —
+  so `wait: true` gates nothing.
+- The vendored DaemonSet is a **formatting-only** transform of upstream, proven rather than
+  asserted: `yq -P 'sort_keys(..)'` of the generated file and of the raw upstream file `diff`
+  clean, so passing it through kustomize changed no content. The committed body's sha256 matches
+  the one recorded in its own provenance header (`80ebd070…`), so the file has not drifted from
+  what was generated.
+
+Gate (pinned toolbox, from this worktree, with the `ENGINE` shim bind-mounting the main checkout's
+gitdir and `DEV_KUBECONFIG=/dev/null`): see the PR description for the per-stage table. **`rc 3`
+(UNDETERMINED), with exactly the three expected skips** — `spec-immutability` (sibling spec repo is
+not beside a worktree), `gitea-settings` (no CI token), `conformance` (no cluster in CI). Every
+other stage passed. **rc 3 is not green** and is not reported as such.
+
+## What I did NOT verify
+
+**Nothing about this component's runtime behaviour. It has never run.**
+
+- **The plugin has never executed.** No pod has been scheduled, no image has been pulled, no scan
+  has happened. It cannot run until the operator merges this and Flux reconciles — `kubectl apply`
+  and `flux reconcile` are operator actions (CLAUDE.md §5), so this is expected, not an oversight.
+- **The resource has never been advertised.** `gpu.intel.com/i915` does not exist on any node and
+  I have not seen it exist. The claim that it *will* be `i915` rather than `xe` is derived from
+  source plus the established fact that `i915` is bound — it is a derivation, not an observation.
+- **The DaemonSet has never been admitted.** That the Kyverno policies do not match this namespace
+  is read off the policy manifests; no admission request has actually been evaluated. Nor has the
+  `kubectl kustomize` output ever been server-side dry-run against the API — rendering is not
+  admission.
+- **The `open()` test has NOT been re-run.** The `EPERM` is exactly as it was. Nothing in this PR
+  touches Plex, so Plex is still software-transcoding and will continue to until the later renderer
+  PR lands. The decisive evidence is an `open()` as **uid 1000 through `s6-setuidgid plex`**;
+  passing as uid 0 is weaker and must not be accepted as proof.
+- **A green gate proves the manifests render, nothing more.** And this gate was not green: it was
+  `rc 3`. Note also that the Gitea `gate` job runs `./ci/gate.sh || [ "$?" -eq 3 ]`
+  (`.gitea/workflows/gate.yml:138`), so a green tick on this PR cannot distinguish rc 0 from rc 3,
+  and `vars.CLUSTER_AVAILABLE` is unset so no L2/L3 conformance runs in CI.
+- **`seLinuxOptions: type: container_device_plugin_t`** is carried verbatim from upstream. These
+  nodes are Debian 13 and I did not measure whether SELinux is enabled or whether that type exists
+  there. Expected to be inert (runc skips labelling when SELinux is disabled), but unverified — if
+  the pods fail to start with an SELinux error, that line is the first suspect.
+- **Renovate has not been run** against this pin. The inline-marker idiom is copied from
+  `platform/kubevirt` / `platform/external-snapshotter`, but whether Renovate normalises the
+  `v`-prefix mismatch (release `v0.37.0` vs image `0.37.0`) for this dep is unknown. If it proposes
+  `newTag: v0.38.0`, the fix is an `extractVersion` packageRule in `renovate.json`.
+- **Resource limits unexamined under load.** The upstream 115Mi/100m limits are taken as-is; I have
+  no measurement of this plugin's footprint on these nodes. Worth a glance given
+  `keycloak-operator` is OOMing at 91% of its limit elsewhere in this cluster.
+- **`clusters/dev` renders it, and there is no dev cluster** (the operator confirmed the
+  `k3s-dev-*` nodes are gone). That is inert rather than wrong — the unified graph (R-ENV-02) has
+  both overlays render the same base — but it is not a thing I could test.
+- **Branch protection / required-checks state not re-measured** in this session. CLAUDE.md §2b
+  says re-measure before relying on it; I did not, so treat the approval/gating claims in §2 as
+  unconfirmed-by-me.
@@ -0,0 +1,482 @@
+---
+date: 2026-10-01
+branch: feat/oauth2-proxy-api-route-per-app
+title: The config model had no way to say "this app's API lives here" — so --api-route could not ship. Now it can, per app.
+---
+
+## What changed
+
+Second half of PR #28 (`decisions/2026-09-30-oauth2-proxy-cookie-collision-fleet.md`), which
+fixed the cookie collision, the 5-minute session and the missing metrics listener in one renderer
+function and **deliberately declined `--api-route`** for a reason I have preserved rather than
+worked around: that flag was measured on **one** app, and `PiHole serves no /api/ at all`, so a
+renderer-wide constant would apply one app's path shape to five. The `auth` block is
+`additionalProperties: false`, so there was no per-app override hook. **Creating the hook was the
+work; setting the flag is the easy part.**
+
+**Three surfaces, in order of what they protect:**
+
+| Surface | Change | What it guards that the others cannot |
+|---|---|---|
+| `schemas/servicedefinition.schema.json` (**1.8.0 → 1.9.0**) | new optional `auth.apiRoutes`: array of anchored path regexes, `uniqueItems`, default `[]` | rejects a too-broad regex **where a human types it** — `ci/l0.sh` validates `catalog/*.yaml` against this schema on every gate run |
+| `catalog/render.py` | emits one `--api-route=<r>` per declared entry, in order; **raises** on an unanchored entry | `ci/render-golden.sh` runs on every gate invocation even when the schema stage skips |
+| `conformance/tests/l0-static/oauth2-proxy-api-routes-per-app/` (new) | cross-checks each rendered sidecar against its own ServiceDefinition | catches a **renderer constant**, which neither of the above can see |
+
+**The `auth` block stays `additionalProperties: false`.** One named, documented key was added; the
+door is not opened. That property is the reason the gap existed, and it is worth keeping — it is
+what made "there is no override hook" a *fact* rather than a *guess* for the two engineers who hit
+it before me.
+
+**Rebased mid-flight.** `origin/main` advanced from `08deb20` to `cdf3e14` while I was working
+(PR #27 and PR #32, the `hardware.gpu: intel-quicksync` capability). **That change also took schema
+`1.8.0`**, so mine is renumbered to **1.9.0** and chained behind it; both version entries are
+present in the description, and `catalog/render.py` had a two-sided conflict in the constants block
+which is resolved by keeping both (the `DRI_*` constants and `OAUTH2_API_ROUTE_PATTERN` are
+unrelated). Re-rendering after the rebase still changes **only** jackett and radarr — plex's new
+golden reproduces byte-identically, so the two changes do not interact.
+
+**Shape:** a **list** of regexes, not a single string, because Radarr needs two
+(`^/api/` and `^/signalr/`) and `--api-route` is natively repeatable. A single string would have
+forced alternation into the value, and alternation is exactly how breadth gets smuggled past a
+prefix-only reading of a regex.
+
+**The anchoring ratchet.** Each entry must match `^\^/[A-Za-z0-9._~/-]+(\.\*|\$)?$` — anchored at
+`^/`, then literal path characters, with an optional trailing `.*` or `$`. The same pattern is in
+the schema, in the renderer, and in the new check, and assertion (5) of the check **fails if they
+drift apart**, because if they do the loosest one wins silently. Rejected shapes, all five
+provoked: `^/`, `^/.*`, `.*`, `/api/` (unanchored), `^/api/|^/` (alternation). Alternation is
+**refused rather than supported**: the flag is repeatable, so one entry per route costs nothing.
+
+**Why breadth is the dangerous direction, and it is not symmetric.** A *missing* `--api-route` is a
+silent 302 — bad, invisible, recoverable. A *too broad* one answers a plain **page load** with 401:
+a logged-out operator gets a bare error page and is **never offered a sign-in**, so the service
+looks permanently broken and no reload recovers it. That is a worse outcome than the defect the
+flag fixes, which is why it is gated on three surfaces rather than documented in one.
+
+### Also folded in: the selftest could exit 1 with ZERO diagnostics
+
+Not my change originally — Desmond Oyelaran found it while wiring `catalog/render.py --selftest`
+into the `render-golden` stage (his PR #33, which touches only `ci/render-golden.sh` and is **not**
+in my base; `origin/main` at `cdf3e14` does not yet drive the selftest from any gate stage). The
+coordinator asked me to fold the fix into this PR because the defect is in the file I own.
+
+`_selftest()` reached `_cookie_name()` through a bare `next(genexpr)` with **no default**. Delete
+the `--cookie-name` emission — exactly the mutation that assertion exists to catch — and it raises
+`StopIteration`, which escapes `_selftest()` **before its collected `SELFTEST FAIL:` lines are
+printed**. Right verdict, no diagnostics, and it masks every other assertion the same mutation
+broke. Fixed with `next(..., None)` plus an explicit `check()` so a missing flag reports **as a
+named assertion**.
+
+Provoking it turned up **three more instances of the same failure class** in the same block, which
+is why this is not literally a one-line change: `side = fa_containers[1]`,
+`side["env"][0][...]`, and the `fa_svc` lookup all raise rather than assert when the sidecar is
+absent entirely. All four are now None-tolerant, and the two fixtures I added for `--api-route`
+were written the same way from the start. Measured, on throwaway copies:
+
+| Mutation | Before | After |
+|---|---|---|
+| `--cookie-name` emission deleted (origin/main's renderer) | rc 1, **0** `SELFTEST FAIL` lines, 1 `StopIteration` | rc 1, **3** `SELFTEST FAIL` lines, 0 tracebacks |
+| the oauth2-proxy sidecar not appended at all | rc 1, **0** lines, `IndexError` | rc 1, **20** `SELFTEST FAIL` lines, 0 tracebacks |
+
+Note the consequence of his stage change for anyone reading this later: once PR #33 lands,
+`.gitea/workflows/gate.yml:138` runs `./ci/gate.sh || [ "$?" -eq 3 ]`, so **rc 1 is not tolerated
+in CI the way rc 3 is** — a selftest failure will block. Which is the right outcome, and it is why
+a selftest that prints nothing was worth fixing before that lands rather than after.
+
+### Declined, with reasons: the metrics-ingress NetworkPolicy
+
+The coordinator asked for my judgement on bundling Ines Dubois's finding — that #28's metrics
+listener genuinely serves (she measured `http=200`, 10,208 bytes,
+`oauth2_proxy_requests_total{code="302"} 2` via port-forward) but the scrape will be **dropped**,
+because namespace `jackett`'s default-deny has **no `ingress` key at all** and
+`jackett-allow-gateway-ingress` admits `fromEntities: [ingress]` only. The fix belongs in each
+app's own rendered policy, i.e. in my file and my goldens. **I declined to bundle it**, and asked
+for it as the immediate next PR. Five reasons, in the order that decided it:
+
+1. **The real reason is that it is a MODEL question, not a policy to bolt on.** Schema 1.6.0
+   already owns this: `observability.metricsEndpoint` renders the Service port **and** the
+   monitoring-namespace ingress allow. What the model cannot currently express is a **second**
+   metrics endpoint on the same pod belonging to a *sidecar the renderer injected*, rather than to
+   the workload the operator declared. Keying a bespoke policy off the pod label
+   `homelab.keithjackson.dev/auth-mode: forward-auth` works today, but it hardcodes "forward-auth
+   implies a metrics port on 44180" into a policy instead of **deriving** it from the sidecar
+   declaration that is already the single source of that port — the exact "list, don't derive"
+   anti-pattern `oauth2-proxy-cookie-names-distinct` assertion (5) exists to forbid, and which I
+   would be introducing one directory away from the check that forbids it. That deserves its own
+   review, and it is squarely the kind of decision I was brought in to make deliberately.
+2. **Different subject, and it is a trust-boundary change.** Mine is the auth *response code* on
+   API paths (R-IDP-02). That is an ingress allow (R-NET-01/02, R-OBS-01) admitting a new
+   namespace to a new port. Bundling a security-relevant ingress rule into a PR whose headline is
+   `--api-route` is how such a rule gets waved through on the strength of the other half — and this
+   PR already needs a human for a *different* sensitive-path reason, so the second subject would be
+   reviewed in the shadow of the first.
+3. **The conflict argument does not actually apply to my diff.** It assumes I am regenerating five
+   goldens; **I am regenerating two** (jackett, radarr), because the flag renders only where
+   declared. A metrics-ingress policy renders a *different document* — a CiliumNetworkPolicy — in
+   all five. Different hunks, in files git merges independently; and goldens regenerate
+   deterministically from the renderer, so the follow-up rebases onto `main` cleanly. The N-1
+   rebase problem `decisions/README.md` records was two branches appending to **the same hunk**,
+   which this is not. I checked my own diff before answering rather than accepting the premise.
+4. **Every fact in the finding is hers and I have measured none of it** — not the 200/10,208 bytes,
+   not the `wget: download timed out` on `:44180` *and* `:4180`, not the Flux sibling ordering that
+   makes `oauth2-proxy` and `apps-wave-0` unordered. Shipping a security-adjacent rule on wholly
+   inherited evidence, as a footnote in someone else's PR, is precisely what CLAUDE.md §2b exists
+   to stop. In its own PR I would measure it first.
+5. **Sequencing: the PodMonitor is in HER PR #29, not mine.** If I ship the policy and her
+   PodMonitor is reworked or delayed, I have opened an ingress allow for a scraper that selects
+   nothing — a hole with no consumer. The policy is only useful paired with the PodMonitor, which
+   argues for it landing adjacent to that, not here. It also drags in a
+   `KNOWN_EXTERNAL_MONITORS` entry in `conformance/tests/l0-static/servicemonitor-ports-resolve/`,
+   a second conformance interaction on top of the new check this PR already adds.
+
+Diff size is the tiebreaker rather than the argument: **1,245 insertions across 13 files**,
+including a ~400-line new conformance check. I would rather hand a reviewer one subject they can
+actually falsify. **Queued, not dropped** — and if the next person takes it, the model note in (1)
+is the part worth arguing about, not the YAML.
+
+### Which apps got which routes, and the measurement behind each
+
+| App | `auth.apiRoutes` | Evidence |
+|---|---|---|
+| **jackett** | `^/api/` | 2026-09-30, its own access log: **33 × 302 against 2 × 401** on `/api/` paths, and **4 of 6 POSTs answered by the proxy and never reaching Jackett** (upstream field `-`, not `/`), including both `/api/v2.0/server/update` calls |
+| **radarr** | `^/api/`, `^/signalr/` | `^/signalr/`: measured **by me, on the live post-fix pod** — see below. `^/api/`: measured on radarr by the PR #30 author, on the **pre-fix** pod |
+| **pihole** | *(none — deliberate)* | serves no `/api/`; **0** unauthenticated requests in its current pod's 23 h |
+| **sonarr** | *(none — not measured)* | 13,643 log lines, **0** unauthenticated requests. No traffic, not a negative result |
+| **qbittorrent** | *(none — not measured)* | 13,649 log lines, **0** unauthenticated requests. Same |
+
+**`^/signalr/` on radarr is the live proof this change was needed**, and it is the strongest
+evidence in this entry. 23 h after PR #28 deployed, on `radarr-7d55cbffb9-42hmh` (0 restarts), the
+sidecar logged **232 unauthenticated requests across the pod's entire lifetime, and all 232 are the
+same call**: `POST /signalr/messages/negotiate`, browser UA, **302**, upstream `-` (Radarr never saw
+one of them). That is a **left-open browser tab** whose session the cookie rename orphaned, polling
+SignalR forever, silently redirected every time, and **never re-prompting the operator**. It is the
+reported symptom — "reloading never fixes it and I am never re-prompted" — still happening in
+production, on a service nobody was looking at. With 401 the UI's connection breaks **loudly**; the
+operator reloads; `/` matches neither regex, so the reload signs them back in.
+
+**`^/api/` on radarr: included, and here is the honest basis.** The current pod has logged **zero**
+`/api/` traffic, so I did **not** re-measure it. It rests on a measurement of **radarr** recorded in
+the merged PR #30 handoff — its 156 unauthenticated requests were `/api/v3/command` (62),
+`/api/v3/queue/details` (46) and `/signalr/messages/negotiate` (43), i.e. **108 of 156 were
+`/api/v3` XHRs**. That is this app's own evidence, by another engineer, not an analogy to another
+app. I include it because `/api/v3/` is a pure JSON namespace a browser never navigates to (Radarr
+serves its UI at `/`, assets under `/Content/` and `/MediaCover/`), so the breadth risk is nil and
+the measured benefit is 70% of that app's rejection volume.
+
+**Sonarr and qBittorrent were measured and deliberately left alone.** Sonarr is nearly the same
+codebase as Radarr, which is exactly what makes `^/api/` + `^/signalr/` a tempting guess and still a
+guess. I read both pods' logs: **zero** unauthenticated requests, nothing but health probes, because
+nobody has opened either UI since the cookie fix rolled. **That is evidence of no traffic, not
+evidence of no need.** Each catalog file now carries a comment saying so, with the numbers and what
+to measure before adding the routes.
+
+### Goldens and fixtures regenerated
+
+Rendered with the canonical command
+(`python3 catalog/render.py --all catalog -o apps --wave-layout --no-image-automation`).
+**Two goldens changed, not five** — and that is the per-app property proving itself on the real
+tree rather than in a fixture:
+
+    apps/wave-0/jackett.yaml   + --api-route=^/api/
+    apps/wave-1/radarr.yaml    + --api-route=^/api/
+                               + --api-route=^/signalr/
+
+`apps/wave-0/pihole.yaml`, `apps/wave-0/qbittorrent.yaml` and `apps/wave-1/sonarr.yaml` are
+**byte-identical to `main`**. **No L1 fixture needed regenerating**: all three
+`conformance/tests/l1-rendering/*/fixtures/svcdef*.yaml` declare `auth.mode: forward-auth` with no
+`apiRoutes`, so they render no flag and their goldens are unchanged. PR #28 touched 5 goldens and 3
+L1 fixtures because it changed flags on *every* sidecar; this one changes flags only where a
+measurement says so, which is the whole point.
+
+### The L0 decision (and it is a decision, not a default)
+
+**Not** in `oauth2-proxy-cookie-names-distinct`. That check's subject is cookie **identity** — its
+inputs are names and domains, its failure mode is mutual session eviction, its register is the
+shared proxy's claim on the default name. `--api-route` shares none of those. Bundling them would
+make that check's name a lie and couple two unrelated ratchets, so a future api-route decision could
+only be expressed by weakening a *cookie* check. That is PR #28's own stated discipline, applied to
+my change rather than quietly dropped.
+
+**A sibling check, `oauth2-proxy-api-routes-per-app`** (whole-repo scope, so it runs on every gate
+invocation via `ci/l0.sh`'s `--include-scope` union — no bespoke stage, which
+`whole-repo-invariants-wired` would fail). **Nowhere was the wrong answer for exactly one reason:**
+the existing gates cannot see a renderer constant. `render-golden` compares the render to the
+goldens, so a constant **plus regenerated goldens agree with each other while both being wrong** —
+I provoked that and it regenerates with **rc 0**. The new check's assertion (2) compares each
+rendered sidecar against its **own ServiceDefinition**, which is the only reading that fails it.
+
+## What I verified
+
+- **`--api-route` is SINGULAR, read off the PINNED binary rather than taken from the brief.**
+  `podman run --rm quay.io/oauth2-proxy/oauth2-proxy:v7.8.1 --help`:
+  `--api-route strings   return HTTP 401 instead of redirecting to authentication server if token
+  is not valid. Format: path_regex`. There is **no `--api-routes`** in the help output. The PM asked
+  me to verify rather than trust, and that is the result — the singular form is confirmed and the
+  help text also confirms the *semantics* I am relying on (401 instead of a redirect).
+- **The premise, in-cluster, read-only, through the pinned toolbox, `--context k3s-prod` named
+  explicitly** (CLAUDE.md §2b and §5). PR #28 **is** deployed: all five sidecars carry
+  `--cookie-name=_oauth2_proxy_<app>`, `--cookie-domain=<app>.home.keithjackson.dev`,
+  `--cookie-refresh=2m30s` and `--metrics-address=0.0.0.0:44180`. Pods, all **0 restarts**:
+  `jackett-6877fcd9f5-ht9d9`, `pihole-…`, `qbittorrent-5547bdb9bd-zq557`,
+  `radarr-7d55cbffb9-42hmh`, `sonarr-ddf485775-8ckwt` (started 2026-09-30T21:41–21:42Z).
+- **The full post-fix log census, parsed rather than grepped** (oauth2-proxy's access-log format,
+  with `/ping` health probes excluded), over each pod's whole life:
+
+      app            lines   unauth  AuthSuccess  removing session  cookie sig invalid
+      jackett       13,649        0            0                 0                   0
+      pihole        13,649        0            0                 0                   0
+      qbittorrent   13,649        0            0                 0                   0
+      radarr        14,099      232            0                 0                   0
+      sonarr        13,643        0            0                 0                   0
+
+  Three findings, kept separate:
+  1. **The collision fix works.** Jackett was 39 unauthenticated before; it is **0** in 13,612
+     non-probe lines. The eviction fingerprint (`removing session` / `cookie signature not valid`)
+     is **0 on all five**, where radarr previously had 8 and 8.
+  2. **Radarr's 232 are one request.** 100% `POST /signalr/messages/negotiate`, 100% **302**, 100%
+     upstream `-`, browser UA (`Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537…`). **Zero `/api/`
+     traffic of any kind** on this pod.
+  3. **0 `AuthSuccess` anywhere**, so nobody has signed in since the restart and `--cookie-refresh`
+     is **still** unexercised — unchanged from PR #28's own caveat.
+- **Only two goldens change, and three are byte-identical to `main`.** `git status` after the
+  canonical render lists exactly `apps/wave-0/jackett.yaml` and `apps/wave-1/radarr.yaml`.
+  `render-golden` is green.
+- **`catalog/render.py --selftest`: PASS**, with new assertions that (a) the `fa` fixture, which
+  declares nothing, renders **no** `--api-route`; (b) a fixture declaring two routes renders exactly
+  those two, in order; (c) the renderer **refuses** all five breadth shapes; (d) four legitimate
+  routes (`^/api/`, `^/signalr/`, `^/api/v3/.*`, `^/api/v2.0/server/update$`) are **accepted** — a
+  ratchet that blocks real API prefixes is a ratchet someone eventually deletes.
+- **The new check: rc 0, 43 assertions, 6 sidecars audited**, with its inventory printed so the
+  reviewer can see what it read:
+  `jackett -> ['^/api/'] | pihole -> (none) | qbittorrent -> (none) | radarr -> ['^/api/',
+  '^/signalr/'] | sonarr -> (none) | crafty -> (none)`.
+- **The check was tested in BOTH directions, on throwaway copies outside the repo**, never by
+  mutating the branch. Each mutation fired its own assertion class and nothing else:
+  - **Pre-change tree** (`git archive origin/main` + the test dropped in): **rc 1, 11 failures** —
+    no service declares routes, the renderer emits none, the renderer does not read `auth.apiRoutes`,
+    the pattern is absent from both renderer and schema, and the shipped renderer accepts all five
+    breadth shapes.
+  - **pihole's golden hand-edited** to carry a route it never declared: **rc 1, 2 failures** (the
+    catalog↔golden mismatch, and "declares nothing must render nothing").
+  - **radarr's golden widened** to `--api-route=^/`: **rc 1, 2 failures** (unanchored, plus the
+    mismatch).
+  - **Renderer turned into a constant and EVERY golden regenerated to match: the regen exits 0** —
+    i.e. `render-golden` is **happy** — and the new check exits **1 with 15 failures**, naming
+    pihole, qbittorrent, radarr and sonarr individually. **This is the case that justifies the
+    check's existence** and I provoked it rather than argued it.
+  - **Schema `items.pattern` deleted**: **rc 1, 1 failure** — the three-surface drift guard.
+  - **Unmutated control: rc 0.**
+- **The selftest diagnostics fix was PROVOKED, not reasoned.** Both mutations above were run
+  against throwaway copies of the renderer outside the repo (`origin/main`'s `render.py` for the
+  "before" column, mine for the "after"), never by mutating the branch. The pre-fix StopIteration
+  and the post-fix 3-and-20 named assertions are quoted from those runs. `--selftest` on the
+  unmutated branch: **PASS**. Goldens after the fix: **unchanged** (it touches only `_selftest()`).
+- **The check is selected by the whole-repo union.** `./conformance/run.sh --layer l0 --requirements
+  r-nosuchreq-99 --include-scope whole-repo --print-selection` lists
+  `tests/l0-static/oauth2-proxy-api-routes-per-app` among **21** tests (it was 20), i.e. it runs
+  even when the touched-requirement filter matches nothing.
+- **THE SCHEMA GUARD IS PROVEN WITH THE PINNED VALIDATOR `ci/l0.sh` ACTUALLY DRIVES**, not by
+  reading the pattern. `./schemas/validate.sh` (`check-jsonschema, version 0.29.4`) against seven
+  throwaway ServiceDefinitions written outside the repo:
+
+      good            ["^/api/", "^/signalr/"]   rc 0
+      bad-root        ["^/"]                     rc 1  apiRoutes ... does not match
+      bad-dotstar     ["^/.*"]                   rc 1  apiRoutes ... does not match
+      bad-unanchored  ["/api/"]                  rc 1  apiRoutes ... does not match
+      bad-alt         ["^/api/|^/"]              rc 1  apiRoutes ... does not match
+      bad-everything  [".*"]                     rc 1  apiRoutes ... does not match
+      bad-mode        ["^/api/"] on mode: oidc   rc 1  const 'forward-auth'
+
+  So a too-broad regex is rejected **where a human types it**, on every gate run, by the real tool
+  — and the shape invariant fires too.
+- **THE RENDERER'S REFUSAL REACHES `render-golden` AS A READABLE FAILURE, NOT A TRACEBACK.** With
+  `^/signalr/` swapped for `^/` in a *copy* of the catalog (outside the repo), both invocation
+  shapes exit **rc 1 with 0 tracebacks** — the single-SD form and `--all <dir> -o <dir>
+  --wave-layout --no-image-automation`, which is the one `ci/render-golden.sh` drives:
+
+      render.py: radarr: auth.apiRoutes entry '^/' is not an anchored path regex. Each entry
+      must start with '^/' and continue with literal path characters (e.g. '^/api/',
+      '^/signalr/', '^/api/v3/.*'). ... A too-broad --api-route answers a PAGE LOAD with 401, so
+      a logged-out operator sees a bare error page and is never offered a sign-in: the service
+      looks permanently broken rather than merely unauthenticated.
+
+  So the breadth guard is gated by `render-golden` even on a run where the schema stage skipped,
+  and the operator is told *why* rather than handed a stack trace.
+- **The schema's new shape invariant.** A non-empty `auth.apiRoutes` requires
+  `auth.mode: forward-auth`, because the flag renders onto the per-app sidecar and no sidecar exists
+  for `oidc`/`native`/`none` — declaring it elsewhere would be a **silent no-op**, which is the
+  failure class this repo keeps paying for.
+- **Gate:** toolbox `./ci/gate.sh` with the two-mount ENGINE shim — exit code and stage list in the
+  next section, because it is not a pass and that belongs there.
+- **No mutating cluster access.** `kubectl get/logs/jsonpath` only, `--context k3s-prod` named on
+  every call. No `apply`/`edit`/`patch`, no `flux reconcile`. No Secret `.data` or `.stringData` was
+  read or printed; I did not need to touch a Secret at all this session. The one container I ran
+  locally was the pinned oauth2-proxy image, to read its `--help`.
+- **`~/.git-credentials` was never read.** PR creation used `git credential fill`, which does not
+  echo the value.
+
+## What I did NOT verify
+
+- **NO BROWSER ROUND-TRIP. BY ANYONE. FOR THIS FLAG.** This is the most important line here. The
+  302/401 split, the dropped POST bodies and the orphaned SignalR poll are all read from access
+  logs; that a **401 repairs** the operator's experience is reasoned from HTTP semantics, not
+  watched. Nobody has added a Jackett indexer or recovered a Radarr tab through this
+  configuration. The human test below is the first real evidence this change will ever have.
+- **Local gate: rc 3 (UNDETERMINED), not 0.** Green: `spec-immutability`, `decisions-log`, `lint`,
+  `l0`, `render-golden`, `l1`. **Skipped:** `gitea-settings` (no `SETTINGS_API_TOKEN` locally) and
+  **L2/L3 conformance** (`CLUSTER_AVAILABLE` unset). Nothing FAILED; nothing runtime was certified
+  either. **rc 3 is UNDETERMINED, never green.**
+- **MY LOCAL `approval-policy: green` WAS MEANINGLESS, AND I ALMOST REPORTED IT AS EVIDENCE.**
+  `ci/approval-policy.sh:47` reads `GITHUB_EVENT_NAME`, which is unset outside CI, so the stage
+  short-circuits at line 50: `ok "not a pull_request event (event='none') — approval policy N/A"`.
+  **A clean local gate therefore CANNOT catch a sensitive-path problem — structurally, not by
+  accident.** Flagged to me by the coordinator from Ines Dubois's round, and I verified it by
+  re-running the gate with `GITHUB_EVENT_NAME=pull_request PR_NUMBER=34`: the stage then
+  **evaluates** and fails closed on the missing token —
+
+      [fail] no SETTINGS_API_TOKEN — cannot read this PR's changed files or its reviews.
+      [fail] stage approval-policy: FAIL
+      [fail] approval-policy: a sensitive change lacks a non-author approval -> blocks merge
+      [fail] MERGE GATE: BLOCKED        (rc 1, not 3)
+
+  — while **every other stage stays green** (`spec-immutability`, `decisions-log`, `lint`, `l0`,
+  `render-golden`, `l1`). That is the useful result: it bounds the CI failure to this one stage
+  rather than leaving "something is red" as the whole report. Fail-closed is correct here and
+  deliberate: unlike `gitea-settings`, where a missing credential is a lock-out, a skip of this
+  stage would be a bypass (CLAUDE.md §1).
+- **CI, read from each commit's `/statuses` rather than from a tick.** Both contexts report, and
+  **only one of them runs**. Identical on `99560c2d`, `df1b25b6` and `f1d16f44`:
+
+      gate / gate (pull_request)         state=failure   desc='Failing after ~1m43-1m45s'
+      gate / conformance (pull_request)  state=success   desc='Has been skipped'
+
+  **That second line is §2b's trap in the wild**: a *skipped* Gitea job reports
+  `state=success`. The conformance job did not execute, so its green tick certifies nothing —
+  read the description, not the colour. The duration is given as a range on purpose: it differs
+  by a second or two per run, and pinning one number inside the file whose commit triggers the
+  next run is a self-reference, not a measurement. The load-bearing fields are the **state** and
+  the **description**, and those are stable across all three.
+
+  Worth recording that the **same context changed its answer over time**: on the first push it
+  briefly read `state=pending desc='Blocked by required conditions'` — still waiting on the gate
+  job it depends on — and resolved to `success` only once that job settled. So sampling a status
+  once can show either a pending **or** a green tick for a job that never executed.
+  **I could NOT read the CI job log to confirm the failing stage from CI's
+  own output**: this Gitea's `/api/v1/repos/{owner}/{repo}/actions/runs` endpoint answers
+  **404** (only `/actions/tasks` exists, which lists tasks without logs), and
+  `/actions/jobs/{task_id}/logs` answers `404 Not found` for the task id. So my attribution of the
+  CI failure to `approval-policy` rests on the local event-set reproduction above plus the
+  sensitive-path rule, **not** on CI's own log. That is weaker evidence than PR #28 had and I am
+  naming it rather than borrowing its wording.
+- **Branch protection and the runner, RE-MEASURED this session against the live API** (§2b),
+  not inherited from CLAUDE.md §2's table:
+
+      GET /repos/Home-Network/home-network-k3s-impl/branches/main
+        protected              True
+        user_can_push          False
+        enable_status_check    True
+        status_check_contexts  ['gate / gate (pull_request)']
+        required_approvals     1
+      GET /admin/actions/runners
+        id=1 name=incluster-dind status=online
+
+  So `required_approvals=1` applies to **every** PR here regardless of path — this needs a human
+  even without the sensitive-path rule. I did **not** provoke a rejected push to `main` this
+  session; `user_can_push: False` is the API's word, which is exactly the weaker form of evidence
+  CLAUDE.md §2 warns about, and I did not strengthen it.
+- **`^/api/` ON RADARR IS NOT MY MEASUREMENT.** It is PR #30's, taken on the **pre-fix** pod. The
+  current pod has logged **zero** `/api/` traffic in 23 h, so I could not reproduce it. I judge the
+  breadth risk nil (`/api/v3/` is pure JSON; the UI is at `/`), but if someone wants a strictly
+  self-measured configuration, drop `^/api/` and keep `^/signalr/` — the 232 are entirely SignalR.
+- **SONARR AND QBITTORRENT ARE UNMEASURED, NOT CLEARED.** Zero traffic in 23 h is zero evidence.
+  qBittorrent's WebUI drives everything through `/api/v2/` and is *plausibly* affected the same way
+  Jackett is; Sonarr is nearly Radarr's codebase. **Both are left alone on purpose** — plausibility
+  is what put a renderer constant on the table in the first place. Expect to add `^/api/` to
+  qBittorrent and `^/api/` + `^/signalr/` to Sonarr once someone opens those UIs and records the
+  split. **This PR does not fix them.**
+- **PiHole is asserted to serve no `/api/` on inherited evidence, not mine.** Its log shows 0
+  unauthenticated requests, which says nothing about what paths it serves. The claim comes from
+  PR #28/#30. Note the pinned `update.range` is `>=2024.0.0`, and PiHole v6 *does* introduce a REST
+  API at `/api/` — so if PiHole is ever upgraded across that line, this needs re-measuring. It is
+  still correct to declare nothing today, because declaring an unmeasured route is the mistake.
+- **`--cookie-refresh=2m30s` has STILL never been exercised** — 0 `AuthSuccess` on all five pods.
+  PR #28 named this as the fix most likely to need a second pass and it remains unobserved. My
+  change does not touch it, but a sign-in during the human test below will exercise both at once,
+  so a failure there must be attributed carefully.
+- **The regexes are not validated against the apps' real route tables.** I assert `/api/v3/` and
+  `/signalr/` are JSON/RPC namespaces a browser never navigates to from Radarr's and Jackett's
+  observed traffic and from how those UIs are built — **not** from reading their routing code or
+  enumerating their endpoints. A path I have not thought of that matches `^/api/` and *is*
+  browser-navigated would 401.
+- **`OPTIONS` preflight and non-browser clients are unconsidered.** `--skip-auth-preflight` exists
+  and is not set; nothing in these logs shows a CORS preflight, and Sonarr/Radarr talk to Jackett
+  **directly on `:9117`**, bypassing the proxy entirely — so I do not expect impact. Unverified.
+- **The live pods do not carry this change and nothing re-measures that they will.** git is not the
+  cluster (§2b). **Merging is what deploys**, and it will roll **jackett** and **radarr** only
+  (`strategy: Recreate` on RWO Longhorn volumes, so a brief gap per app, and radarr/sonarr lose
+  direct `:9117` access to Jackett while its pod restarts). Existing sessions survive — no cookie
+  flag changed — so unlike PR #28 nobody is signed out.
+- **No L2/L3 assertion reads any of this.** There is no runtime check that a `/api/` request to a
+  gated app answers 401, and I did not write one; `CLUSTER_AVAILABLE` is unset so it could not run
+  in CI anyway. The whole guard is static.
+- **`catalog/render.py --selftest` is UNGATED IN MY BASE, and I did not wire it.** Desmond's PR #33
+  does that (`ci/render-golden.sh`) and is in review, not merged — `origin/main` at `cdf3e14`
+  contains no stage that drives it, which I checked rather than assumed (`grep -n selftest
+  ci/render-golden.sh` → no match). **So until #33 lands, my new selftest assertions are
+  dev-time only**, and the *gated* half of my breadth ratchet is the schema `pattern` (via
+  `ci/l0.sh`'s schema stage), the renderer's own `raise` (via `render-golden`), and the new L0
+  check. I deliberately did not wire the selftest myself: #33 already does it, and duplicating it
+  here would collide with his branch in the one file he owns. Note the two PRs are **independent
+  but complementary** — his makes the stage real, mine makes its output readable when it fires.
+- **I did NOT verify PR #33's own claims.** The 106-assertion count, the
+  selftest→stage→`gate.sh`→`BLOCKED` propagation chain, and the `--cookie-refresh 2m30s → 10m`
+  mutation being caught by the selftest but not by `render-golden` are all his measurements,
+  relayed to me by the coordinator. I reproduced only the StopIteration defect itself, because that
+  is the part I was asked to fix. If his branch is reworked, nothing in mine depends on it.
+- **The new check does not read the spec repo** and asserts nothing about `../home-network-k3s`
+  (R-TEST-06, read-only). `schemas/renderer-contract.md` is updated in **this** repo only; the
+  canonical spec `docs/13-service-catalog.md` is **not** updated, because it is immutable from here.
+  Someone with spec write access should mirror `auth.apiRoutes` into docs/13 and docs/15.
+- **THE OAUTH2-PROXY SIDECARS' METRICS ARE STILL UNSCRAPABLE, AND THIS PR DOES NOT FIX IT.** See
+  the declined-bundle section above. #28 shipped `--metrics-address` as a prerequisite and said
+  plainly that it was not scraped; Ines has now measured that the listener **serves** and that the
+  scrape is **dropped by each app namespace's default-deny**. Nothing in my change touches that.
+  Anyone reading this entry as "the fleet gates are observable now" is wrong, and the next PR is
+  where that changes.
+- **Alternation is refused, and that is a judgement I may have to revisit.** No app in this catalog
+  needs it, and the flag is repeatable — but if some future app genuinely needs a regex the ratchet
+  rejects, the fix is to widen the pattern **with a reason** on all three surfaces at once, not to
+  delete the ratchet on one.
+- **CI status: two contexts, and only one of them ran.** Quoted in the PR body and in the report to
+  the PM. `approval-policy` **fails by design** until a non-author approving review exists, because
+  the diff touches `conformance/**`, which `infra/gitea/approval-policy.json` lists as sensitive
+  ("the tests and the harness; changing them changes what green MEANS"). That is CLAUDE.md §1
+  working, not a defect. `required_approvals=1` applies to every PR here regardless of path, so this
+  needs a human either way.
+
+### The human test — and how to tell a too-broad regex from a working one
+
+After merge, wait for **new** pod names at 2/2 in `jackett` and `radarr`.
+
+1. **The breadth test FIRST, before anything else, from a signed-out browser** (private window).
+   Open `https://radarr.home.keithjackson.dev/` and `https://jackett.home.keithjackson.dev/UI/Dashboard`.
+   **Each must bounce through Keycloak and land on the app's UI.** If either shows a bare
+   `401 Unauthorized` instead of a sign-in, **the regex is too broad — stop and revert.** That is
+   the one failure this change can introduce, and it is the only test that can see it. A working
+   regex is invisible on a page load; a broken one is total and immediate.
+2. **Jackett, the original report:** signed in, *+ Add indexer* → pick one → *Configure* → *Okay*.
+   **It must save.** Previously 4 of 6 POSTs never reached Jackett at all.
+3. **Radarr, the live symptom:** open Radarr and leave the tab open. DevTools → Network, filter
+   `signalr`. Today those negotiate calls are **302** and the UI never complains. They must now be
+   **200** while signed in — and if the session is ever lost, **401**, with the UI showing a visible
+   connection error instead of polling silently forever.
+4. **The loud-failure test:** in DevTools, delete the `_oauth2_proxy_radarr` cookie and let the tab
+   poll. The negotiate call must answer **401** and Radarr must *show* a problem. Then **reload** —
+   `/` matches neither regex, so the reload must sign you back in silently. That round trip is the
+   entire fix: fail loudly on the API, recover on the page load.
+5. **The six-minute test (PR #28's open item, still open):** stay on a dashboard for >6 minutes,
+   then save something. That exercises `--cookie-refresh=2m30s` past the 300 s access-token cliff.
+   If it fails, suspect the realm's refresh/SSO-idle timeouts — not this change.
+6. **If something still fails, it must now fail LOUDLY.** A silent success-looking no-op means the
+   change did not take: check the running pod's args for `--api-route` before assuming anything
+   else.
@@ -0,0 +1,211 @@
+---
+date: 2026-10-01
+branch: feat/quicksync-via-device-plugin
+title: "`hardware.gpu: intel-quicksync` is now satisfied by `gpu.intel.com/i915`, not a `/dev/dri` hostPath — the hostPath was measured not to work (device-cgroup EPERM), and the swap took ZERO ServiceDefinition edits"
+---
+
+## What changed
+
+The mechanism behind `hardware.gpu: intel-quicksync` was replaced. **No ServiceDefinition field
+moved** — that was the whole design bet when this field was made an enum of capability classes
+instead of a device path, and this is the change that collects on it.
+
+**Why the old mechanism had to go.** The `/dev/dri` hostPath shipped on 2026-09-30 was measured
+**non-functional** in the live Plex pod on k3s-prod-2 on 2026-10-01. The device node was *visible*
+in the container — `crw-rw---- root video2 226,128 renderD128` — and could not be opened:
+
+| attempt | result |
+|---|---|
+| `uid 0`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+| `uid 1000`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+
+`EPERM` is the **device cgroup**, not `EACCES` from file permissions, proven three ways: root on the
+**host** opens the same device `rc=0`; the node is unified `cgroup2fs`; and the container's runc
+`config.json` read `linux.resources.devices: [{allow: False, access: 'rwm'}]` with
+`linux.devices: None` — one deny-all rule, no allow entry for major 226. A hostPath is a *filesystem*
+operation: it binds the device node in and adds **no cgroup allow rule**, because Kubernetes cannot
+infer one from a volume. `uid 0` failing matters: root does not bypass the cgroup device gate, so
+"run it as root" was never going to work either, and `privileged: true` is unavailable (Kyverno
+`restrict-host-namespaces` is Enforce on app namespaces).
+
+**`catalog/render.py`** — retired `DRI_VOLUME_NAME` / `DRI_DEVICE_DIR` for `QUICKSYNC_RESOURCE`
+(`gpu.intel.com/i915`). The `hardware.gpu == intel-quicksync` branch now emits exactly one thing on
+the **app container only** (never a sidecar):
+
+```yaml
+resources:
+  limits:
+    gpu.intel.com/i915: 1
+```
+
+and no volume, no volumeMount, no `securityContext`, no `nodeSelector`. `limits` rather than
+`requests` is mandatory, not stylistic: Kubernetes mirrors `requests` from `limits` automatically and
+a requests-only extended resource is invalid. This is the **only** app-container `resources` render
+path in the file and it is deliberately narrow — no general CPU/memory resources feature was invented
+along the way.
+
+**Self-tests** — the three existing ones updated plus one new:
+- the emitted `resources` dict is asserted **exactly** (`{"limits": {"gpu.intel.com/i915": 1}}`), so
+  a stray CPU/memory addition through this field fails loudly;
+- the retired hostPath volume and `/dev/dri` mount are asserted **absent**;
+- the no-`hardware` negative extended to cover `resources` (absence still renders nothing at all);
+- the `podSecurityContext` pass-through still holds, and the "no synthesised `securityContext`"
+  assertion is kept verbatim;
+- **new:** the i915 limit must never reach the oauth2-proxy sidecar.
+
+**`schemas/renderer-contract.md`** — the mechanism half of `### hardware` and invariant 8 rewritten.
+The intent half is unchanged and still correct. The `EPERM` measurement is recorded as the *reason*
+the design is what it is, so a future reader cannot "simplify" it back without knowing it was tried
+and measured. The `DirectoryOrCreate` rationale is kept, explicitly labelled historical.
+
+**`catalog/plex.yaml`** — comment correction only; the `hardware:` field is byte-identical.
+`git diff` on that file contains **no non-comment lines**. Its old comment restated the mechanism
+("exactly two things — a `/dev/dri` hostPath volume…"), which is now false; it states intent and
+points at the contract, and keeps the measured facts that remain true (device on all three nodes,
+`i915` bound, the cont-init hook / `supplementalGroups` explanation).
+
+**`runbooks/plex-quicksync.md`** — the diagnostic sequence now checks
+`.status.allocatable["gpu.intel.com/i915"]` and the plugin DaemonSet instead of a hostPath; the
+decisive open test is specified **as uid 1000 via `s6-setuidgid plex`**; the cgroup `EPERM` is
+recast from a *risk* to the *reason for the design*; a new top-of-page "STOP HERE IF PLEX IS
+`Pending`" section and a "Plex will not schedule" procedure carry the new failure mode and the escape
+hatch. The manual-toggle section (`HardwareAcceleratedCodecs`, Settings → Transcoder, still not
+settable by any env var) is unchanged.
+
+**`apps/wave-0/plex.yaml`** — re-rendered: six hostPath lines removed, a three-line `resources` block
+added. Nothing else in `apps/` changed.
+
+### The one real regression, stated rather than sold
+
+This trades failure modes, and the new one is **worse in blast radius**:
+
+| | device/resource unavailable ⇒ |
+|---|---|
+| **Before** | `DirectoryOrCreate` mounted an empty dir; Plex ran and **silently software-transcoded**. Invisible, harmless to everything else. |
+| **After** | Plex is **Unschedulable**. `apps-wave-0` is `wait: true` and `apps-wave-1` `dependsOn` it, so this **stalls the app tier**. |
+
+It is **unavoidable**: an extended resource is a hard scheduling constraint, and it is also the only
+thing that produces the cgroup allow rule. There is no graceful-degradation variant. Plex's
+schedulability is now coupled to the `intel-gpu-plugin` DaemonSet — broken, removed, or merely not
+yet reconciled, it takes the app tier with it. That is the price of hardware transcoding working at
+all.
+
+**Escape hatch — one field, no code change:** delete the `hardware:` block from `catalog/plex.yaml`
+and re-render. Plex returns to software transcoding and the tier unblocks. It is in the contract, in
+the runbook's "Plex will not schedule → remedy B" with exact commands, and here, because it is what
+someone will need at 2am.
+
+## What I verified
+
+**The precondition, re-read from the branch rather than assumed:** the plugin is a standalone
+DaemonSet at `platform/intel-gpu-plugin/` with its own Flux Kustomization
+(`clusters/base/intel-gpu-plugin.yaml`, named `intel-gpu-plugin` in `flux-system`), namespace
+`intel-gpu-plugin`, DaemonSet `intel-gpu-plugin`, pinned `intel/intel-gpu-plugin:0.37.0`. Every
+`kubectl` command I wrote into the runbook names those exact objects. The node/DaemonSet readings
+themselves (`i915=1` on all three nodes; 3 desired / 3 ready / 3 available) were **measured by the
+PM, not by me** — see "What I did NOT verify".
+
+**`render.py --selftest`: PASS**, and every new assertion is **mutation-tested — 8/8 mutants caught**
+(a test that cannot fail is not a test):
+
+| mutant | caught by |
+|---|---|
+| `requests` instead of `limits` | exact-resources assertion |
+| emission dropped entirely | exact-resources assertion |
+| `/dev/dri` hostPath + mount re-added | hostPath-absent assertion |
+| emitted unconditionally (no `hardware` key) | no-`hardware` negative |
+| `supplementalGroups` synthesised | no-pod-`securityContext` assertion |
+| resource renamed `gpu.intel.com/xe` | exact-resources assertion |
+| quantity as the string `"1"` | exact-resources assertion |
+| limit leaked onto the oauth2-proxy sidecar | new app-container-only assertion |
+
+**Re-render is clean and scoped.** `python3 catalog/render.py --all catalog -o apps --wave-layout
+--no-image-automation` then `git status --porcelain apps/` ⇒ `M apps/wave-0/plex.yaml` and nothing
+else. `git diff --stat apps/` ⇒ `1 file changed, 3 insertions(+), 6 deletions(-)` — exactly the six
+hostPath lines out, the three-line `resources` block in. **`apps/wave-1/keycloak.yaml` and
+`apps/wave-0/vm-demo.yaml` are untouched** (checked explicitly: they are git-tracked but excluded
+from the golden comparison by `RENDER_GOLDEN_NOT_ONBOARDED` in `ci/render-golden.sh`, so a spurious
+diff there would land ungated). `clusters/` is untouched; `--emit-waves` was not run.
+
+**Invariant 8 holds, mechanically checked rather than eyeballed.** `git diff -U0 catalog/plex.yaml`
+filtered to non-comment changed lines returns **empty**, and `yaml.safe_load` gives
+`hardware == {'gpu': 'intel-quicksync'}`. Zero ServiceDefinition field edits.
+
+**The gate: `rc 3` UNDETERMINED, which is this environment's calibrated best — NOT green, and I am
+not calling it green.** Run from this worktree in the pinned toolbox
+(`DEV_KUBECONFIG=/dev/null`, plus an `ENGINE` shim bind-mounting the main checkout's `.git`
+read-only, per CLAUDE.md §3; `infra/ansible/.cache/prod.kubeconfig` was **not** copied):
+
+| stage | result |
+|---|---|
+| `spec-immutability` | **SKIPPED** — spec repo not present at `/work/../home-network-k3s` (R-TEST-06 unverified) |
+| `decisions-log` | green |
+| `lint` | green |
+| `l0` | green |
+| `render-golden` | green |
+| `l1` | green |
+| `approval-policy` | green (`event='none'`, N/A off a PR event) |
+| `gitea-settings` | **SKIPPED** — `SETTINGS_API_TOKEN` unset |
+| `conformance` | **SKIPPED** — `CLUSTER_AVAILABLE != true`, L2/L3 not run |
+
+Three skips, the three expected ones; **nothing failed**. `diff base: origin/main @ 8214642d87de …
+HEAD @ 78bf68493d7d`, touched requirements resolved (`R-PLAT-02` among them), so the R-TEST-03
+coverage gate actually applied — an earlier run on the uncommitted tree reported the touched set
+"NOT DETERMINABLE (on-base)" and that is *not* the run I am quoting.
+
+## What I did NOT verify
+
+**HARDWARE TRANSCODING IS STILL NOT OBSERVED WORKING, and nothing in this branch could establish
+it.** A green gate proves the **manifest renders**, nothing more.
+
+- **The `open()` test has not been re-run under the new mechanism, and cannot be until this merges
+  and Flux reconciles.** The whole case for this change is that the device plugin's
+  `DeviceSpec{… Permissions: "rw"}` produces the cgroup allow rule the hostPath could not. That is
+  read from the v0.37.0 **source**, not observed on this cluster. If the plugin does not in fact
+  deliver it here, this change buys nothing and costs schedulability.
+- **The Plex-side toggle is still unset.** `HardwareAcceleratedCodecs` was **absent** from
+  `Preferences.xml` as of 2026-09-30 and nothing in this branch touches it; no env var, ConfigMap or
+  `postBuild` substitution can. It is a manual UI click, and until it happens Plex software-transcodes
+  no matter how the device arrives.
+- **I did not measure the cluster at all.** `i915=1` on all three nodes and the DaemonSet's
+  3/3/3 come from the PM's brief, dated 2026-10-01. Per CLAUDE.md §2b this repo is not evidence about
+  the live platform, and neither is a handoff entry: **re-measure before relying on them** (the
+  commands are first-command #1 in the runbook). The DaemonSet was 2m15s old when read — young enough
+  that it had not been observed stable for any length of time.
+- **The new failure mode has not been exercised.** "Plugin unavailable ⇒ Plex Unschedulable ⇒
+  `apps-wave-0` stalls ⇒ `apps-wave-1` blocked" is derived from `wait: true` and `dependsOn` in the
+  Flux wiring, not from a drill. The escape hatch (delete `hardware:`, re-render) is likewise
+  reasoned, not rehearsed — though it is a strict subset of the renderer path the selftest covers.
+- **No L2/L3 conformance.** `CLUSTER_AVAILABLE` is unset, so the gate's runtime layer did not run
+  here and does not run in CI either. `spec-immutability` and `gitea-settings` also skipped.
+- **CI's own tick, when it appears, will not distinguish rc 0 from rc 3**:
+  `.gitea/workflows/gate.yml:138` is `./ci/gate.sh || [ "$?" -eq 3 ]`. Read the stage output, not the
+  colour — and take the newest row per context from `GET /statuses/{sha}`, which appends.
+- **`imagePullPolicy: Always` on `plexinc/pms-docker:latest` is still a live hazard** to the
+  cont-init-hook reasoning this feature leans on. Unchanged by this branch, recorded in the runbook's
+  floating-image caveat.
+
+### Post-merge sequence for the operator (all mutating steps are yours, not an agent's)
+
+1. Merge the PR once the gate reports. Confirm Flux applied it — compare
+   `kubectl --context k3s-prod -n flux-system get kustomization apps-wave-0
+   -o jsonpath='{.status.lastAppliedRevision}'` against the merge commit.
+2. Confirm Plex **scheduled** (this is the new risk):
+   `kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex -o wide`.
+   `Pending` ⇒ runbook "Plex will not schedule".
+3. Confirm it consumed the resource:
+   `kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex
+   -o jsonpath='{.items[0].spec.containers[0].resources.limits}'`.
+4. **The decisive test — as uid 1000, via `s6-setuidgid plex`. `rc=0` is the pass condition:**
+   `kubectl --context k3s-prod -n plex exec deploy/plex -- s6-setuidgid plex sh -c
+   'exec 3<>/dev/dri/renderD128 && echo RW-OPEN-OK || echo RW-OPEN-DENIED'`.
+   **Passing as `uid 0` is weaker evidence and must not be accepted as proof** — Plex runs as 1000,
+   and uid 0 was one of the two `EPERM`s that killed the hostPath.
+5. **Only then** the Plex UI toggle: Settings → Transcoder → "Use hardware acceleration when
+   available" (under *Show Advanced*) → Save. Then force a transcode and read Dashboard → Activity:
+   `Transcode (hw)` is the proof; plain `Transcode` is software.
+
+Step 4 failing means the plugin is not delivering the cgroup allow rule on this cluster. Do **not**
+reach for `privileged: true` (Kyverno Enforce rejects it, and uid 0 failed anyway) and do **not**
+re-add a `/dev/dri` hostPath (that is the mechanism this replaced, measured non-functional). The
+escape hatch is to delete the `hardware:` block from `catalog/plex.yaml` and re-render.
@@ -0,0 +1,205 @@
+---
+date: 2026-10-01
+branch: fix/render-selftest-runs-in-gate
+title: catalog/render.py's 106 assertions were dev-time only; the gate now runs them and blocks on them
+---
+
+## What changed
+
+Single subject, one file: **`ci/render-golden.sh` now invokes `catalog/render.py --selftest`**
+and exits 1 when it fails. Nothing else. No renderer change, no new conformance test, no new
+gate stage.
+
+This closes the top follow-up of `decisions/2026-09-30-oauth2-proxy-cookie-collision-fleet.md`:
+
+> Nothing in `ci/` or `.gitea/` runs `catalog/render.py --selftest` — I grepped for it. So the
+> twenty-odd assertions in that selftest, including the six I just added, are dev-time only: a
+> future edit that drops either flag *and* re-renders the goldens would pass the entire merge
+> gate.
+
+Re-measured here and confirmed: before this commit, `grep -rl -- --selftest` over `ci/` and
+`.gitea/` returned **nothing** for the renderer. The selftest executes **106** assertions
+(`check()` invocations counted at runtime via `sys.settrace`, without editing `render.py`; 109
+static call sites, three of them on failure-only branches). All 106 were unexercised by the
+merge gate.
+
+### The lead was WRONG, and here is why
+
+The brief named `conformance/tests/l0-static/auditability-proofs-producible/check.sh:291` as the
+one-line fix. That line is real and it does drive selftests:
+
+    for tool in ("record_run.py", "status.py"):
+        r = run([sys.executable, str(LIB / tool), "--selftest"])
+
+But it is **the wrong hook**. That loop is section (d) of a check that serves **R-SEC-03**, and
+its subject is the *evidence chain* — can a self-skipped run, or a stale ledger row, become an
+audit proof? `record_run.py` and `status.py` are **in** that chain; `catalog/render.py` is not.
+Hanging the renderer there would have made the check's name, its `intent` annotation and its
+requirement mapping (`R-SEC-03`) all untrue, and a future decision about the renderer could then
+only be expressed by weakening a *security-auditability* check. That is the precise coupling the
+cookie-collision entry refused for its own check, and it applies here.
+
+The second option that entry offered — a new scope-declaring L0 test — is the supported mechanism
+for a **whole-repo invariant** (`ci/l0.sh` Stage 3b's `--include-scope whole-repo` union, ratcheted
+by `tests/l0-static/whole-repo-invariants-wired`). It is the right shape for an invariant
+quantified over the tracked tree. It is the wrong shape for *"a script in this repo must pass its
+own unit test"*: it would need to declare requirements it does not serve, and it would be two new
+files of several hundred lines by this tree's conventions for what is a six-line invocation.
+
+**The honest home is `ci/render-golden.sh`** — the gate stage whose subject *already is*
+`catalog/render.py`. It already resolves `$PYTHON` and `$RENDER_PY`, already reports rc 3 when
+either is missing, already exits 1 on renderer error, and `ci/gate.sh` already maps it through
+`for s in lint l0 render-golden l1` where a FAIL is BLOCKED and a SKIP is UNDETERMINED. The
+stage's two assertions are independent and both are needed:
+
+- `render-golden` catches golden-vs-renderer drift, but is satisfied by **any** renderer whose
+  output was re-rendered — including one that just lost a contract invariant and had all ten
+  goldens regenerated in the same commit. That is exactly the hole named in the quote above.
+- `--selftest` catches that, and is blind to golden drift.
+
+The file header and the stage name were updated so neither is a lie: the stage is now "the
+renderer is trustworthy", with two numbered assertions, and the step summary emits its own
+`- render-selftest:` line so the gate output still NAMES what ran.
+
+The selftest runs **before** the empty-catalog `exit 3` on purpose: it renders synthetic
+ServiceDefinitions in memory and is catalog-independent, so an empty catalog must not be able to
+skip it.
+
+`catalog/render.py` itself was **not touched** (owned by a concurrent branch this session), and
+neither was any file under `conformance/`.
+
+## What I verified
+
+- **The selftest runs and passes: rc 0, `render.py selftest: PASS`, 106 assertions executed.**
+  Counted at runtime, not from the source: a `sys.settrace` hook counting `check()` calls, in a
+  throwaway script outside the repo, with no edit to `render.py`.
+- **The stage passes with it wired in**, host run of the single stage:
+
+      [gate] Stage: render==golden reproducibility (canonical render vs committed apps/)
+      [gate] Stage: catalog/render.py --selftest (the renderer's own output-contract assertions)
+      [ ok ] render.py --selftest passed
+      [ ok ] render==golden: all 10 onboarded service golden(s) reproduce from the canonical render
+
+  rc **0**. The golden comparison still reports all 10 goldens, so the new invocation did not
+  displace what the stage already asserted.
+- **IT CAN FAIL — provoked twice, against a mutated COPY of the renderer, never against the
+  tree.** `ci/render-golden.sh` already honours `RENDER_PY`, so each mutant lived in
+  `/tmp/claude-1000/desmond-selftest/` and the repo's `catalog/render.py` was never modified —
+  which matters, because that file belongs to a concurrent branch this session.
+  - **Mutation 1 — the regression this selftest was written for.** Deleted the
+    `f"--cookie-name=_oauth2_proxy_{s['name']}"` emission from `_oauth2_sidecar_container()`,
+    i.e. reverted the fleet-wide cookie-collision fix of PR #28. Stage rc **1**. Note the shape
+    of the failure, because it is a real (minor) weakness and is listed as a follow-up: the
+    selftest died with an uncaught `StopIteration` out of `next(a for a in container["args"] if
+    a.startswith("--cookie-name="))` at `render.py:2641`, rather than printing its collected
+    `SELFTEST FAIL:` lines. It is correctly non-zero and the stage correctly blocks, but the
+    operator gets a traceback instead of the assertion message.
+  - **Mutation 2 — a clean, named assertion failure.** `OAUTH2_COOKIE_REFRESH = "2m30s"` ->
+    `"10m"`, widening the refresh past the measured 300s access-token lifetime. Stage rc **1**:
+
+        [fail] catalog/render.py --selftest FAILED — the renderer violates its own output contract.
+        [fail]   Every golden under apps/ is produced by this renderer, so the goldens below cannot
+        [fail]   be trusted even if they reproduce byte-for-byte. Reproduce with:
+        [fail]     python3 catalog/render.py --selftest
+            SELFTEST FAIL: --cookie-refresh=10m parses to 600.0s, which is not inside (0, 300s). A
+            refresh at or past the access-token lifetime never renews in time, which is the same
+            logout it was added to stop
+
+    This is the mutation that proves the wiring's value most directly: mutation 2 would **not**
+    be caught by `render-golden`'s diff if the goldens were regenerated in the same commit, and
+    it is one of the two flags the cookie-collision entry explicitly recorded as **ungated**.
+- **THE NON-ZERO PROPAGATES ALL THE WAY TO `gate.sh`'s rc — provoked, in the pinned toolbox, not
+  reasoned about.** `RENDER_PY=/desmond-selftest/render-mutant2.py ./ci/gate.sh`:
+
+      [ ok ] stage spec-immutability: pass
+      [ ok ] stage decisions-log: pass
+      [ ok ] stage lint: pass
+      [ ok ] stage l0: pass
+      [gate] Stage: catalog/render.py --selftest (the renderer's own output-contract assertions)
+      [fail] catalog/render.py --selftest FAILED — the renderer violates its own output contract.
+          SELFTEST FAIL: --cookie-refresh=10m parses to 600.0s, which is not inside (0, 300s). ...
+      [fail] stage render-golden: FAIL
+      [ ok ] stage l1: pass
+      ...
+      [fail] render-golden: FAILED -> blocks merge
+      [fail] MERGE GATE: BLOCKED
+
+  `gate.sh` rc **1**. Three links in the chain, each measured: selftest rc 1 -> stage rc 1 ->
+  gate rc 1 / `MERGE GATE: BLOCKED`. This mattered enough to provoke rather than assume: this
+  repo has a documented habit of exit codes getting flattened, and
+  `.gitea/workflows/gate.yml:138` runs `./ci/gate.sh || [ "$?" -eq 3 ]`, which tolerates rc 3 —
+  so a stage that degraded a failure to a skip would be swallowed by CI. rc 1 is not tolerated
+  there, and rc 1 is what this produces.
+- **`spec-immutability` genuinely RAN and passed, it did not skip.** The two-mount worktree shim
+  plus `-e SPEC_REPO=/home-network-k3s` (the fix recorded in the 2026-09-30 cookie entry) works:
+  `stage spec-immutability: pass`, not SKIPPED. R-TEST-06 is therefore actually covered by these
+  runs.
+- **No collision with PR #29.** The brief warned that `feat/oauth2-proxy-auth-failure-visible`
+  had added checks under `conformance/tests/l0-static/`. Measured —
+  `git diff origin/main...origin/feat/oauth2-proxy-auth-failure-visible --name-only` — it touches
+  **no `conformance/` path at all**: `decisions/`, `platform/oauth2-proxy/` (a promtool test, an
+  alert rule, its kustomization), `runbooks/`. Disjoint from this branch either way, since this
+  branch touches only `ci/render-golden.sh` and this file.
+- **Priya's files were not touched.** `catalog/render.py`, `schemas/servicedefinition.schema.json`,
+  `apps/wave-*` and `conformance/tests/l1-rendering/` are all untouched by this branch — confirmed
+  by `git diff --name-only`, which lists exactly two paths.
+- **`catalog/render.py --selftest` was the LAST unwired selftest in the tree.** Measured:
+  `grep -rl -- --selftest` across the repo finds selftest-capable artifacts in exactly four
+  places, and after this commit every one is driven by the gate — `conformance/lib/record_run.py`
+  and `conformance/lib/status.py` by `auditability-proofs-producible/check.sh:291`, the Keycloak
+  client-secret reconciler (embedded in `platform/keycloak/client-secret-reconciler.yaml`) by
+  `keycloak-client-secret-reconciler/check.sh:633`, and `catalog/render.py` by this change.
+
+## What I did NOT verify
+
+- **Local gate: rc 3 (UNDETERMINED), NOT 0. This is not green and must not be reported as
+  green.** Measured in the pinned toolbox on `eeb6211`:
+
+      [ ok ] spec-immutability: green
+      [ ok ] approval-policy: green
+      [ ok ] decisions-log: green
+      [ ok ] lint: green
+      [ ok ] l0: green
+      [ ok ] render-golden: green
+      [ ok ] l1: green
+      [fail] MERGE GATE: UNDETERMINED — greenness was NOT ESTABLISHED.
+      [fail]   Stages that did not run: gitea-settings
+
+  Nothing FAILED. The single cause is `gitea-settings` **SKIPPED** — `SETTINGS_API_TOKEN` is not
+  set in my environment, so branch protection and runner registration were not read; CI has that
+  token and the stage is expected to run there. **L2/L3 `conformance`** also skipped
+  (`CLUSTER_AVAILABLE` unset) but is a *declared* gap that `gate.sh` treats as not-applicable
+  rather than undetermined, so it is not what produced the 3. Touched requirements
+  **R-GITOPS-01 R-TEST-02 R-TEST-03 R-TEST-06** are therefore **UNVERIFIED at L2/L3**, all four.
+  The stage this branch changes, `render-golden`, is green **and** was independently proven able
+  to turn the whole gate red (above).
+- **`approval-policy` passed LOCALLY only because there is no pull_request event**
+  (`not a pull_request event (event='none') — approval policy N/A`). That is not evidence that it
+  passes. This diff touches `ci/**`, a sensitive path, so in CI the stage is expected to **FAIL**
+  until a non-author approving review exists. That is §1 working, fail-closed, not a defect.
+- **The selftest's own quality is unchanged and ungated.** I wired its *invocation*; I did not
+  review its 106 assertions, and nothing ratchets them. Deleting the body of `_selftest()` would
+  still pass this stage, because an empty selftest returns 0. The repo already has the right
+  answer for that — `conformance/tests/l0-static/keycloak-client-secret-reconciler/mutations.py`
+  injects regressions one at a time and requires `--selftest` to catch every one — and applying
+  that pattern to `catalog/render.py` is the follow-up, deliberately not done here: it is a
+  second subject, it touches `conformance/`, and it would need to live beside the renderer whose
+  file another branch owns this session. **Until it exists, "the selftest runs" is proven and
+  "the selftest is falsifiable" rests on my two hand-run mutations, which are not in the gate.**
+- **Mutation 1's traceback is not fixed.** `render.py:2641`'s `next(...)` has no default, so the
+  absence of `--cookie-name` raises `StopIteration` and the selftest's collected `SELFTEST FAIL:`
+  lines are never printed. Correct verdict, poor diagnostics, and it suppresses the other
+  assertions that mutation also broke. The one-line fix (`next(..., None)` plus a `check()`) is in
+  `catalog/render.py`, which is another branch's file this session, so it is a follow-up and not
+  a thing I touched.
+- **Nothing about the live cluster.** No `kubectl` of any kind was run on this branch — not even
+  read-only; this change is offline CI wiring and needed none. No `apply`/`edit`/`patch`, no
+  `flux reconcile`. Merging this deploys nothing: it changes only what the gate refuses.
+- **CI's own run of this branch.** Quoted in the PR body from the commit's `/statuses`, with
+  descriptions read rather than colours. A *skipped* Gitea job reports `success` with
+  "Has been skipped" and certifies nothing.
+- **`vars.CLUSTER_AVAILABLE` is still unset in CI**, so even a fully green PR proves the L0/L1
+  offline subset and nothing about runtime behaviour.
+- **I did not re-measure branch protection, required approvals, or the runner.** Those are
+  CLAUDE.md §2 claims and §2b says re-measure before relying on them; this branch relies on none
+  of them beyond the PR flow itself, and I read the gate's own output rather than the API.
@@ -0,0 +1,149 @@
+---
+date: 2026-10-02
+branch: fix/plex-quicksync-verification
+title: Hardware transcoding works — and `runbooks/plex-quicksync.md`'s verification section was wrong
+---
+
+## What changed
+
+**Hardware transcoding is confirmed working on prod.** Measured 2026-10-02 from
+`Plex Media Server.log` in the running pod: `vaapi` on `/dev/dri/renderD128`
+(`Intel CoffeeLake-S GT2 [UHD Graphics 630]`), `using hardware decode accelerator vaapi`,
+`zero-copy support present`, `using zero-copy transcoding`, and
+`final decoder: vaapi, final encoder: vaapi`. VAAPI for **both** directions, zero-copy, so frames
+never round-trip through the CPU. The device-plugin mechanism from PR #37 works end to end.
+
+**And `runbooks/plex-quicksync.md`'s verification section was wrong, which is the correction this
+branch exists to publish.** The page said the Plex dashboard showing `Transcode (hw)` was "the *only*
+proof available, and nothing else will ever tell you". The operator had hardware transcoding fully
+working while **Dashboard → Activity showed nothing of the sort** — so the page's authoritative check
+was silent on a working system, and sent someone looking for a fault that did not exist. Three
+inference errors are now corrected on the page, two of them recorded as named false leads so the next
+reader does not re-walk them:
+
+1. **The dashboard is not proof.** Replaced as the primary check by Plex's own log — the grep is in
+   the page, in its **redacting** form, because `Plex Media Server.log` carries API URLs containing
+   `X-Plex-Token` (CLAUDE.md §6; a credential has been leaked into a transcript here before). The
+   page also now says explicitly that **`Plex Transcoder Statistics.log` is the wrong file** — it
+   holds XML `<Variant .../>` records, not ffmpeg output, and the name invites the mistake — and why
+   `grep -a` is required. The dashboard is **kept but demoted** to a secondary, best-effort signal,
+   with its absence stated to prove nothing.
+2. **An fd snapshot is not evidence.** A `/proc/*/fd` sweep inside the pod found no holder of
+   `renderD128` while hardware transcoding was active; transcode sessions are short and the scan
+   landed in a gap. Recorded as a documented false lead, with the reliable process-level check in its
+   place: the transcoder's **argv** (`-hwaccel`, `-hwaccel_device`, `vaapi`, `renderD128`), found via
+   `ps -eo pid,user,comm | grep -i transcoder`.
+3. **A missing preference key does not mean the setting is off.** The page told readers to treat an
+   absent `HardwareAcceleratedCodecs` in `Preferences.xml` as "never enabled". With hardware
+   acceleration demonstrably on and working, that key was **absent from all 27 attributes** while
+   `HardwareDevicePath` **was** written and the log said `hardware transcoding: enabled`. First-command
+   #4 now reads both keys and states plainly that `HardwareDevicePath` present is a positive signal,
+   absence of `HardwareAcceleratedCodecs` means nothing either way, only `="0"` is conclusive, and the
+   log settles it.
+
+Also on the page:
+
+- **GID `992` is no longer presented as a number to expect.** Under the plugin the injected device
+  arrives `root:root`, so the image's `cont-init.d` hook resolves GID **`0`** and adds `plex` to the
+  **root group** — measured `Groups: 0 100 1000`, `id plex` → `groups=1000(plex),0(root),100(users)`.
+  992 was the hostPath-era host `render` GID. The page now says so, marks the root-group membership as
+  an inherent property of the plugin's `DeviceSpec` rather than a defect, and repeats that it is
+  **not** a reason to reintroduce `supplementalGroups` (still a no-op: `s6-setuidgid` rebuilds the
+  group list from `/etc/group`). The stale `Groups: 100 1000` claim in the escalation section is gone.
+- **A new troubleshooting entry for legitimate per-title fallbacks**, marked as a **hypothesis**.
+  Between two fully hardware sessions, one logged `enabled, but no hardware decode accelerator found`
+  and empty final decoder/encoder. UHD 630 decodes H.264, HEVC 8/10-bit and VP9 but **not AV1** and
+  not 10-bit 4:2:2 HEVC, so the likely cause is the source file, not the platform — explicitly
+  written as a lead to check first, **not** a confirmed diagnosis of that item.
+- The status is stated at the top of the page and in a rewritten provenance footer that separates the
+  2026-10-02 confirmations from the superseded 2026-09-30/2026-10-01 hostPath-era readings.
+- `runbooks/README.md`'s index entry for the page was stale in the same two ways (it still described a
+  `/dev/dri` hostPath and the dashboard-only proof) and is corrected to match.
+
+Unchanged on purpose: the "STOP HERE IF PLEX IS `Pending`" section and the escape hatch (delete the
+`hardware:` block from `catalog/plex.yaml`, re-render) keep exactly their existing prominence — they
+are correct and they are the 2am material. The "no alert is possible" reasoning is also kept verbatim
+in substance: Plex declares no `metricsEndpoint`, renders no ServiceMonitor, and a fallback alert
+would be a rule evaluating against nothing.
+
+**Documentation only.** No change to `catalog/`, `schemas/`, `apps/` or `platform/`.
+
+## What I verified
+
+- **Gate, in the pinned toolbox from this worktree** (`DEV_KUBECONFIG=/dev/null`, `ENGINE` shim
+  bind-mounting the main checkout's `.git`): **rc 3 UNDETERMINED**, diff base
+  `origin/main @ 948875aad37d`, with exactly the three expected skips and **no failures**. Green:
+  `decisions-log`, `lint`, `l0`, `render-golden`, `l1`, `approval-policy`. Skipped:
+  `spec-immutability` (spec repo unavailable → R-TEST-06 not verified), `gitea-settings`
+  (`SETTINGS_API_TOKEN` unset), `conformance` (`CLUSTER_AVAILABLE!=true` → L2/L3 not run).
+  Touched requirements: **R-OBS-04**, reported `UNVERIFIED` for lack of a cluster. `approval-policy`
+  passed as `N/A` locally (`event='none'`, not a `pull_request`) — it evaluates for real on the PR.
+  **rc 3 is not green**, it is the calibrated worktree result, and it is reported here as rc 3.
+- **The gate was run three times, with the same rc 3 / three-skip result on each committed tip** —
+  including the final one. Worth recording from the first attempt: run *before* committing, with
+  `HEAD == origin/main`, the touched-requirement set is `NOT DETERMINABLE (on-base)` and the
+  R-TEST-03 coverage gate cannot be applied at all. Commit before gating from a fresh worktree branch.
+- **`runbooks/kustomization.yaml` still carries `- plex-quicksync=plex-quicksync.md`** in the
+  `configMapGenerator` `files:` list (line 110) — checked explicitly, because without it the page
+  merges and still 404s. `README=README.md` is likewise present, which is what serves the corrected
+  index entry.
+- **This change touches no `sensitive_paths`.** `infra/gitea/approval-policy.json` lists `ci/**`,
+  `.gitea/**`, `conformance/**`, `clusters/**`, `bootstrap/**`, `infra/gitea/**`,
+  `platform/kyverno/**`, `platform-config/kustomization.yaml`, `.sops.yaml`, `**/*.sops.yaml`,
+  `**/rbac.yaml`, `CLAUDE.md`. `runbooks/**` and `decisions/**` are not among them, and the
+  `approval-policy` stage passed.
+- **No YAML was added or edited**, so `indent-sequences: consistent` (which blocked a sibling branch
+  today) is not in play here; `lint` passed over the full tracked YAML set regardless.
+- **`DECISIONS-LOG.md` is untouched** — this handoff is its own file, per CLAUDE.md §7.
+- **The PR's own CI run was read, not assumed** (PR #39, run 208, task 261, head `1677dff5688b`).
+  **It is RED, and not because of this branch.** Every stage this change could affect is green —
+  `decisions-log`, `lint`, `l0`, `render-golden`, `l1`, `approval-policy` (`approval-policy` evaluated
+  for real on the `pull_request` event and passed: no sensitive path touched). The single failure is
+  **`gitea-settings`**, quoting the measured field:
+
+      [fail]   required_approvals: DECLARED 1 but LIVE is 0
+      [fail] Gitea has DRIFTED from infra/gitea/expected-branch-protection.json.
+      [fail] MERGE GATE: BLOCKED
+
+  Every other declared setting matched (`block_admin_merge_override`, `block_on_rejected_reviews`,
+  `dismiss_stale_approvals`, `enable_status_check`, `enable_push=false`,
+  `status_check_contexts = ["gate / gate (pull_request)"]`, `runners online: 1`). So **live `main`
+  currently requires ZERO approvals**, the declaration still says 1, and with
+  `REQUIRE_GITEA_SETTINGS: "true"` in `.gitea/workflows/gate.yml` that drift **blocks every PR in
+  the repo**, not just this one. Note the shape of it: the check that would notice approvals are
+  unenforced is the check that is failing. Local branch `fix/admin-approval-override` (`616a497`,
+  "fix(ci): let an administrator merge their own green PR") looks like the in-flight work that means
+  to change the declaration. **I did not touch it** — `infra/gitea/**` is a sensitive path, needs a
+  non-author approval, and is outside this branch's documentation-only scope. It is an operator
+  decision: re-assert the protection, or change the declaration deliberately in its own PR.
+- The repository-wide grep for `992` to find what else still asserts it: remaining occurrences are in
+  `catalog/plex.yaml`, `catalog/render.py`, `schemas/renderer-contract.md`,
+  `platform/intel-gpu-plugin/README.md`, `platform/intel-gpu-plugin/kustomization.yaml`, and three
+  older `decisions/` entries.
+
+## What I did NOT verify
+
+- **I did not re-measure anything on the live cluster in this session.** Every platform number on the
+  page — the VAAPI/zero-copy log lines, `RW-OPEN-OK` as uid 1000, the `gpu.intel.com/i915` limits and
+  requests on the pod spec, `Groups: 0 100 1000`, `id plex`, the 27 `Preferences.xml` attributes, the
+  empty `/proc/*/fd` sweep — came to me as measurements taken earlier on 2026-10-02 and was
+  transcribed, not re-run. Per CLAUDE.md §2b, treat the page as reporting those readings rather than
+  as proof of current state; re-run the primary check before relying on it.
+- **The per-title fallback cause is unverified and labelled as such.** The source codec of the title
+  that logged `no hardware decode accelerator found` was never read. The AV1 / 10-bit-4:2:2-HEVC
+  explanation is inferred from UHD 630's capability set alone.
+- **Whether Plex's dashboard *ever* shows `Transcode (hw)` on this deployment is unknown.** It was
+  observed silent during working hardware transcoding; whether that is a client/version quirk, a
+  reporting bug, or something about this setup was not investigated. The page therefore demotes it
+  rather than explaining it.
+- **The commands as written in the page were not executed from the page.** They are the forms that
+  were run, re-wrapped for `dev.sh`; in particular the `exec -it deploy/plex -- sh` route for the argv
+  check is noted as possibly lacking a TTY under `dev.sh run` and was not exercised in that shape.
+- **Stale `992` copies outside `runbooks/` were left alone**, deliberately — scope was documentation
+  only, and `catalog/plex.yaml`, `catalog/render.py`, `schemas/renderer-contract.md` and
+  `platform/intel-gpu-plugin/` all still carry the hostPath-era `Groups: 44 100 992 1000` reading as
+  if current. `render.py --selftest` asserts around that comment, so correcting it is a real change
+  with gate consequences, not a doc edit. **That is open work.** Older `decisions/` entries also cite
+  992; those are historical records and should stay as written.
+- **CI on the PR.** The branch is pushed and the PR opened; I did not watch the Gitea run report, so
+  nothing here should be read as "CI green". The only gate result I can quote is the local rc 3 above.
@@ -0,0 +1,76 @@
+---
+date: 2026-10-03
+branch: fix/cilium-cgroup-hostroot
+title: Cilium socket-LB never covered host processes — cgroup.hostRoot pointed at an unmounted path
+---
+
+## What changed
+
+`cgroup.hostRoot` changes from `/run/cilium/cgroupv2` to `/sys/fs/cgroup` in all three parity
+copies: `platform/cilium/helmrelease.yaml` (authoritative), `platform/cilium/values.yaml`
+(mirror) and `infra/ansible/roles/cilium/templates/cilium-values.yaml.j2` (bootstrap).
+`autoMount` stays `false`.
+
+**The defect.** With `autoMount: false`, the Cilium chart only bind-mounts whatever exists at
+`hostRoot`, and expects the operator to have mounted cgroup2 there (cilium 1.20.1 values.yaml).
+The repo's rationale was "k3s provides the cgroupv2 mount". That is true only for
+`/sys/fs/cgroup`. Nothing mounts cgroup2 at `/run/cilium/cgroupv2` on these nodes; it is an
+empty `DirectoryOrCreate` hostPath. So the agent mounted cgroup2 there itself, from inside its
+own private cgroup namespace, and that mount's root is the agent container's own leaf cgroup.
+The socket-LB programs (`cil_sock4_connect` etc.) were attached to it and covered only the
+agent.
+
+**The consequence.** Every other host-netns process got no Service/LB translation:
+- Service ClusterIPs were unreachable from every node's host.
+- The node holding the Gateway's L2 lease could not reach the Gateway VIP (`no route to host`).
+  Other nodes reached it only because their packets crossed the wire to the lease holder, where
+  the tc datapath handles them.
+- k3s-prod-1 holds both the Gateway lease and the kube-vip API VIP. Its apiserver could not reach
+  the OIDC issuer, logging `oidc: authenticator not initialized` since at least 2026-09-12
+  (899 occurrences by 2026-10-03). kubectl SSO through the VIP therefore failed, and in-cluster
+  OIDC clients failed on whichever calls landed on that apiserver.
+- Webhooks and aggregated APIs were unaffected only because k3s runs
+  `--enable-aggregator-routing=true` (endpoint IPs, not ClusterIPs).
+
+Cilium's own source (v1.20.1 `bpf/bpf_sock.c`, `__sock4_xlate_fwd`) handles exactly this case
+when the hook does run: a host-netns connect to an L7-LB Service (the Gateway) is rewritten to
+`127.0.0.1:<l7_lb_proxy_port>`, the local Envoy listener. The fix is to make the hook cover the
+host.
+
+## What I verified
+
+- `cilium-dbg status --verbose` on k3s-prod-1: `Socket LB: Enabled`, `Coverage: Hostns-only`.
+  The live cilium-config shows `bpf-lb-sock=false`, but socket LB is implied by kube-proxy
+  replacement. The Gateway Service frontends `192.168.2.20:{80,443}` and
+  `10.43.225.79:{80,443}` map to `127.0.0.1:14337`.
+- Host: `findmnt /run/cilium/cgroupv2` is empty on all three nodes, and `/sys/fs/cgroup` is
+  `cgroup2`.
+- In the agent: `/run/cilium/cgroupv2` is a `cgroup2 none` mount, the agent's own cgroup reads
+  `0::/`, the mount has 0 child cgroups, and `bpftool cgroup show /run/cilium/cgroupv2` lists all
+  `cil_sock*` programs. k3s runs in `0::/system.slice/k3s.service`, and the agent container in
+  `/kubepods.slice/…/cri-containerd-3ad04cde….scope`.
+- **The decisive test**, run on k3s-prod-1 in the same netns: raw TCP connect to
+  `192.168.2.20:443`, `10.43.225.79:443` and `10.43.59.141:8080` CONNECTED from inside the agent
+  container (covered cgroup) and FAILED from the host (`No route to host` / timeout).
+- Blast radius: host-network pods are cilium, cilium-envoy, cilium-operator (API via
+  `192.168.2.9:6443`, not a Service), kube-vip (`127.0.0.1:6443`, `svc_enable=false`) and
+  node-exporter (no outbound). Nothing depends on host Service translation being absent; the
+  change only makes Service/LB addresses reachable from the host where they currently fail.
+- `helm template cilium/cilium --version 1.20.1` with the new values, in the toolbox: the
+  DaemonSet bind-mounts hostPath `/sys/fs/cgroup` into `cilium-agent` and `clean-cilium-state`,
+  the cilium-config shows `cgroup-root: "/sys/fs/cgroup"`, and the init containers are unchanged
+  (no `mount-cgroup`).
+- Toolbox gate: see the PR description.
+
+## What I did NOT verify
+
+- **The rollout itself.** After merge, the agent DaemonSet rolls one node at a time
+  (`maxUnavailable: 1`). On each node, afterwards: host TCP connects to the three targets above
+  must succeed. On k3s-prod-1, the apiserver's background OIDC init retry must stop logging
+  `authenticator not initialized` with no k3s restart.
+- That the old programs, attached to the previous agent container's cgroup, disappear with it.
+  Expected, because cgroup BPF attachments die with the cgroup, but not observed.
+- An end-to-end kubectl OIDC login (needs a browser).
+- No regression test was added. A natural one is an L2 conformance check that a host-netns
+  connect to the Gateway VIP succeeds on the lease holder, but `conformance/**` is a sensitive
+  path. Left as a follow-up.
@@ -0,0 +1,45 @@
+---
+date: 2026-10-03
+branch: fix/node-keycloak-hosts-pin
+title: Pin the OIDC issuer to the Gateway VIP on prod nodes — the WAN answer does not hairpin
+---
+
+## What changed
+
+Follow-up to PR #40 (`decisions/2026-10-03-public-dns-wan-target.md`). Public DNS for
+`*.home.keithjackson.dev` now CNAMEs to the WAN IP. The prod nodes resolve through public DNS,
+and the router does not hairpin, so the kube-apiserver could no longer reach its OIDC issuer
+`keycloak.home.keithjackson.dev`.
+
+- `roles/node_baseline`: a new `node_baseline_hosts_pins` list (default `[]`) and one
+  `blockinfile` task (tags `dns`, `hosts`). It writes the pins into `/etc/hosts` as a single
+  marked block. An empty list removes the block.
+- `group_vars/k3s_prod/vars.yml`: pins `keycloak.{{ k3s_base_domain }}` → `192.168.2.20`, which
+  must stay equal to `GATEWAY_VIP` (parity comment added). It also replaces the stale "no
+  /etc/hosts pin is needed" paragraph with the measured reason the pin is now required.
+
+Dev is unaffected (empty default).
+
+## What I verified
+
+- After PR #40 reconciled (`platform` lastAppliedRevision `576e1ab`, Healthy=True), on
+  k3s-prod-2 and k3s-prod-3: `getent hosts keycloak.home.keithjackson.dev` →
+  `136.51.115.252 wan.home.keithjackson.dev`. `curl https://keycloak…/realms/homelab/.well-known/openid-configuration`
+  → exit 60, `ssl_verify_result=18` (the router's self-signed cert, not the Gateway).
+- `journalctl -u k3s` (last 20 min) showed no OIDC/JWKS errors yet. The apiserver still holds
+  the keys it cached before the flip, so the failure is latent until a k3s restart or key rotation.
+- Neither node has cloud-init (`/etc/cloud` absent), so nothing rewrites `/etc/hosts` at boot.
+- `--check --diff --tags hosts --limit k3s-prod-2,k3s-prod-3`: each node gets exactly the
+  three-line marked block `192.168.2.20	keycloak.home.keithjackson.dev`.
+- Outside-in, PR #40 works: check-host.net probes from 3 external nodes got `200 OK` from
+  `https://plex.home.keithjackson.dev/identity` at `136.51.115.252`.
+
+## What I did NOT verify
+
+- **k3s-prod-1**: Ansible cannot reach it. SSH host key verification fails (`known_hosts:94`,
+  ED25519 `SHA256:zgUS8OeWznzCX9hlN7XdZ8iLNkU1G2loXjtJFfEUYpE`), which is consistent with
+  the 2026-09-11 SSD replacement and reinstall. The operator must confirm the fingerprint
+  before the pin can be applied there.
+- The playbook has not been applied for real. It is an operator run after merge:
+  `ansible-playbook playbooks/node-baseline.yml -e k3s_hosts=k3s_prod --tags hosts`.
+- An end-to-end kubectl OIDC login after the pin (needs a browser).
@@ -0,0 +1,129 @@
+---
+date: 2026-10-03
+branch: feat/plex-remote-access-l4
+title: Plex remote access — raw tcp/32400 via a pinned LoadBalancer VIP, not the L7 Gateway
+---
+
+## What changed
+
+Reported symptom: users cannot reach Plex from outside the LAN. **Two independent causes**, measured
+rather than inferred:
+
+1. **Plex advertised no remote connection at all.** Live `Preferences.xml` on prod:
+   `customConnections` **ABSENT**, `ManualPortMappingMode` **ABSENT**, while
+   `PublishServerOnPlexOnlineKey="1"` — publishing was on and there was nothing reachable to publish.
+   Plex's automatic UPnP/NAT-PMP path cannot work from a pod on the Cilium overlay.
+2. **Nothing from the WAN reached the cluster.** `136.51.115.252:32400` refused;
+   `136.51.115.252:443` answered with a **self-signed router cert** (`C=ZZ, CN=%D`), not the
+   Gateway's LE wildcard. The Plex Service was `ClusterIP` only — no off-cluster path to 32400 by
+   any route.
+
+This branch fixes the in-cluster half. **Remote access does not work until the operator also does
+two manual steps** (`runbooks/plex-remote-access.md`): the WAN tcp/32400 port-forward, and ticking
+"manually specify public port" in the Plex UI.
+
+- **`apps/wave-0/plex-remote.yaml` (new, hand-written).** A `type=LoadBalancer` Service on tcp/32400
+  pinned to `${PLEX_L4_VIP}` via `lbipam.cilium.io/ips`, plus a `CiliumNetworkPolicy`
+  (`plex-allow-remote-32400`) admitting `world`/`cluster`/`host`/`remote-node` to 32400. Non-catalog
+  because `schemas/servicedefinition.schema.json` is `additionalProperties: false` and has no field
+  for a LoadBalancer Service or a pinned VIP — the same reason `apps/wave-0/pihole-dns.yaml`,
+  `platform/crafty/service-l4.yaml` and `apps/vms/unifi/50-service-l4.yaml` are hand-written.
+- **`catalog/render.py`:** an `EXTRA_UNITS` entry (`plex-remote`, anchored to `plex`, `above: False`)
+  so `--emit-waves` places the unit in plex's own wave.
+- **`clusters/base/cluster-vars.yaml`:** new `PLEX_L4_VIP` (dev default `192.168.1.57`), plus the
+  substitution inventory comment updated to name `plex-remote: PLEX_L4_VIP`.
+- **`clusters/prod/kustomization.yaml`:** prod override `PLEX_L4_VIP: 192.168.2.24`.
+- **`apps/wave-0/kustomization.yaml`:** `plex-remote.yaml` added to `resources`.
+- **`runbooks/plex-remote-access.md` (new):** the two manual steps, the four-step verification, why
+  no alert is possible, and the two documented warts below.
+
+**`ADVERTISE_IP` was deliberately NOT set.** Setting it to `https://plex.home.keithjackson.dev:443`
+was the alternative — it needs no new WAN hole and the mechanism was verified working on an
+already-claimed server. **The operator chose the native path (2026-10-03)** so Plex terminates its own
+TLS (`*.plex.direct`) and streams bypass the Envoy L7 hop. `catalog/plex.yaml`'s deferral note for
+`ADVERTISE_IP` remains accurate and untouched.
+
+**`externalTrafficPolicy: Cluster`, with a known cost.** `Local` would preserve the client IP, but
+Cilium's L2 announcer does not follow the backend pod — the wart that cost this LAN an 8-day
+Minecraft outage (see `platform/crafty/service-l4.yaml`). Consequence accepted and documented: SNAT
+puts the client in `192.168.0.0/16`, inside `ALLOWED_NETWORKS`, so **Plex classifies remote streams
+as LAN**. That is a metering effect (remote bandwidth limits will not bind), not an authorization one.
+
+**SENSITIVE PATHS — this PR touches `clusters/**` (two files), so `ci/approval-policy.sh` will BLOCK
+it.** It needs the operator's `/approve-override <reason>` posted as a PR **review** of type
+"Comment". I have deliberately not posted it: CLAUDE.md §1 forbids an agent clearing this on its own
+PR, and the agent account holds `owner`, so the prohibition is a rule, not a limit.
+
+## What I verified
+
+All measurements against live prod, 2026-10-03, read-only.
+
+- **Plex's advertised state**, the finding this branch rests on: `customConnections=<ABSENT>`,
+  `ManualPortMappingMode=<ABSENT>`, `ManualPortMappingPort=<ABSENT>`,
+  `PublishServerOnPlexOnlineKey="1"`, `allowedNetworks="10.0.0.0/8,192.168.0.0/16"`. Read by
+  attribute grep, never by dumping `Preferences.xml` (it holds `PlexOnlineToken`).
+- **The L7 Gateway path works and is NOT the problem.** Via the Gateway VIP,
+  `GET https://plex.home.keithjackson.dev/identity` → `HTTP 200`, `claimed="1"`, real
+  `machineIdentifier`. HTTPRoute `plex-32400`: `Accepted=True`, `ResolvedRefs=True`.
+- **Plex enforces its own auth**, which is what makes an unfiltered WAN port defensible:
+  `/library/sections`, `/status/sessions`, `/:/prefs` all → **401** unauthenticated through the
+  Gateway.
+- **The SNAT/allowedNetworks wart is real, not theoretical.** A tagged probe through the Gateway was
+  logged by Plex as `[10.0.2.214:36040 (Allowed Network (WAN))]` — the Envoy pod IP, classified as an
+  allowed network.
+- **WAN state:** `136.51.115.252:32400` → connection refused. `136.51.115.252:443` → self-signed
+  cert `C=ZZ, ST=Somewhere, O=%D342e7e6b, CN=%D`. Public DNS resolves correctly:
+  `plex.home.keithjackson.dev` → CNAME `wan.home.keithjackson.dev` → `136.51.115.252`.
+- **`192.168.2.24` is free in-cluster.** `CiliumLoadBalancerIPPool default-pool`: IPsTotal 11 /
+  IPsUsed 4 / IPsAvailable 7; the only allocations are `.20` (cilium-gateway-shared), `.21`
+  (unifi-l4-*), `.22` (crafty-l4-*), `.23` (pihole-dns-*). Not a node address (`.10/.11/.12`) and not
+  the kube-vip control-plane VIP (`.9`).
+- **The manual-port-mapping step genuinely cannot be automated.** Grep of the running image's whole
+  init surface (`/etc/cont-init.d`, `/plex-common.sh`, `/usr/local/bin`) for `ManualPortMapping` and
+  `PLEX_PREFERENCE` → **NONE**.
+- **`ADVERTISE_IP` would have worked on a claimed server** (recorded because the runbook's
+  "why not" rests on it): `setPref "customConnections"` at line 126 of `40-plex-first-run` sits
+  **outside** the `PLEX_CLAIM`/`PlexOnlineToken` guard, and the `/.firstRunComplete` early-exit
+  marker is on the **overlay** FS, not the `/config` PVC, so the script re-runs every pod start.
+- **Toolbox gate** (`./infra/devcontainer/dev.sh run bash -c './ci/gate.sh'`): **rc 3**, with
+  `spec-immutability`, `decisions-log`, `lint`, `l0`, `render-golden`, `l1` and `approval-policy` all
+  **pass**. The rc 3 is two environmental skips — `gitea-settings` (no local `SETTINGS_API_TOKEN`)
+  and `conformance` L2/L3 (`CLUSTER_AVAILABLE` unset). **Not green. Reported as rc 3, not as pass.**
+- **`render-golden` passing is the specific proof** that the `render.py` `EXTRA_UNITS` change, the new
+  hand-written unit and the edited wave kustomization are mutually consistent.
+
+## What I did NOT verify
+
+- **THE ACTUAL FIX. Nothing in this branch has been proven to make remote access work** — it cannot
+  be, because both remaining steps are off-repo and manual (router forward, Plex UI toggle). This
+  branch removes two blockers and adds none; it does not demonstrate an outcome.
+- **Outside-in reachability, at all.** Every probe above was taken from inside the LAN, so the
+  `:443` self-signed cert strictly proves only that there is **no NAT hairpin** — it does **not**
+  prove outside-in failure. A router cert on WAN:443 is suggestive (its management UI would shadow a
+  forward), but unproven. **This repo has no off-LAN vantage point.** Test 3 in the runbook
+  (app.plex.tv from a cellular connection) is the only honest check.
+- **`192.168.2.24` was never ARP-probed off-cluster.** Only in-cluster pool accounting was checked.
+  `.23` was `arping`-probed before it was claimed; `.24` was not, so an off-cluster squatter would
+  not have been caught.
+- **The LoadBalancer has never been reconciled.** `plex-l4-tcp` does not exist in prod yet, so
+  LB-IPAM has not actually assigned the pinned VIP, L2 announcement has not been observed for it, and
+  the new CiliumNetworkPolicy has never admitted a real packet. Whether `world`/`host`/`cluster`/
+  `remote-node` is the correct entity set **under this SNAT path** is reasoned from
+  `platform/crafty/networkpolicy.yaml`'s working game-port rule, **not measured**.
+- **Whether Plex's remote path needs anything beyond the manual port tick.** If plex.tv still lists
+  no remote connection after both manual steps, this branch is necessary-but-insufficient.
+- **The CI approval-policy outcome.** The local stage reported green, but that is an offline N/A, not
+  a verdict: this PR touches `clusters/**`, so CI will block it pending an override.
+- **PRE-EXISTING DRIFT I FOUND AND DELIBERATELY LEFT ALONE — worth its own follow-up.** Running the
+  documented `python3 catalog/render.py --emit-waves` regenerates the wave kustomizations and wants to
+  add **`vm-demo.yaml`** to wave 0 and **`keycloak.yaml`** to wave 1. Both are listed as
+  intentionally not onboarded in `ci/render-golden.sh`
+  (`RENDER_GOLDEN_NOT_ONBOARDED="keycloak vm-demo"`); their manifests are tracked in `apps/` but
+  deliberately absent from the `resources` lists so Flux does not apply them. `--emit-waves` places
+  any file present on disk and has **no awareness of that exclusion**, while `render-golden.sh` uses
+  `--wave-layout` and compares rendered manifests only — it never checks kustomization `resources`,
+  which is why this has survived unnoticed. **So the header "DO NOT hand-edit: re-run
+  `render.py --emit-waves`" is currently unsafe advice**: following it onboards an IdP and a KubeVirt
+  VM. I therefore restored both kustomizations and hand-added only my one line. I did not fix
+  `--emit-waves`, and I have not determined how long the drift has existed or whether either service
+  is *meant* to be onboarded.
@@ -0,0 +1,73 @@
+---
+date: 2026-10-03
+branch: feat/public-dns-wan-target
+title: Public DNS now points at the WAN — Gateway-published names CNAME to an in-cluster DDNS record
+---
+
+## What changed
+
+The operator closed the open "public-target gap" recorded in
+`platform/externaldns/helmrelease.yaml`. The WAN NAT / port-forwarding (80/443 → Gateway VIP
+192.168.2.20) was done by the operator outside this repo.
+
+Operator decisions (2026-10-03):
+
+1. **Scope: every Gateway-published name goes public.** That covers the five `exposure: public`
+   apps (plex, gitea, keycloak, nextcloud, seerr) and also the hand-authored admin UIs (auth,
+   crafty, grafana, headlamp, the `home.keithjackson.dev` apex). `internal` services still publish
+   nothing: their `controller: none` route annotation is checked before the target is read.
+2. **The WAN IP is kept current in-cluster**, not hardcoded.
+
+Implementation:
+
+- `platform/gateway/gateway.yaml`: `external-dns.alpha.kubernetes.io/target: wan.${BASE_DOMAIN}` on
+  the shared Gateway. Every gateway-sourced record changes from `A 192.168.2.20` to
+  `CNAME wan.home.keithjackson.dev`.
+- `platform/externaldns/ddns.yaml` (new): `favonia/cloudflare-ddns:1.17.1` Deployment in the
+  `external-dns` namespace. It reuses `cloudflare-api-token` and keeps `wan.${BASE_DOMAIN}` an
+  unproxied IPv4 A record at the WAN IP (detected via Cloudflare trace, checked every 5 min,
+  `DELETE_ON_STOP=false`). external-dns has no TXT registry entry for it, so it never owns or
+  deletes it.
+- The helmrelease comment block and the README now describe the decision instead of the gap.
+
+Known consequence: **k3s nodes resolve through public DNS** (`node_baseline_dns_servers` = 8.8.8.8 /
+1.1.1.1). The kube-apiserver's OIDC issuer `keycloak.home.keithjackson.dev` therefore resolves to the
+WAN IP from the nodes, and kubectl SSO works only if the router hairpins WAN:443 back to the Gateway.
+A LAN-side probe of `136.51.115.252:443` on 2026-10-03 returned a **self-signed placeholder cert**,
+not the Gateway's Let's Encrypt wildcard. That suggests no hairpin. The cert-based admin kubeconfig
+is unaffected.
+
+## What I verified
+
+- Live prod, 2026-10-03, before the change: Gateway `gateway/shared` ADDRESS `192.168.2.20`.
+  External-dns args include `--source=gateway-httproute --source=service --policy=sync
+  --txt-owner-id=prod --domain-filter=keithjackson.dev`, and it logs "All records are already up to
+  date".
+- Public answers at 1.1.1.1: plex/gitea/keycloak/grafana/auth/apex → `192.168.2.20`;
+  `crafty-mc` → `.22` and `unifi` → `.21` (service source, not affected by this change); jackett →
+  none (internal, correctly unpublished).
+- The authority (betty.ns.cloudflare.com) has no unprefixed TXT, AAAA or MX at `plex.` or the apex,
+  so the A→CNAME switch has no co-existing record to collide with.
+- WAN IP from the workstation: `136.51.115.252`.
+- No CiliumNetworkPolicy or NetworkPolicy in `external-dns`. The Kyverno label and default-deny
+  policies match only `app-namespace` namespaces. The vpn egress policy selects only
+  `egress-class: vpn` pods.
+- Toolbox gate result: see the PR description (run after this commit).
+
+## What I did NOT verify
+
+- **The outside-in path.** I have no vantage point outside the LAN. Whether
+  `https://<svc>.home.keithjackson.dev` from the internet reaches the Gateway with the LE cert is
+  UNVERIFIED.
+- **Hairpin / kubectl SSO after the flip.** Likely broken per the probe above. Fix candidates: a
+  node `/etc/hosts` pin for `keycloak.${BASE_DOMAIN}` → Gateway VIP via `node_baseline`, or router
+  NAT reflection.
+- **The record-type flip itself**, until Flux reconciles: whether Cloudflare accepts external-dns's
+  create-CNAME in the same batch as the delete-A, or converges one interval later.
+- `favonia/cloudflare-ddns` under this pod securityContext. The README warns that
+  `no-new-privileges` misbehaves on some Proxmox-style hosts. These are bare-metal nodes, but it is
+  untested until the pod runs.
+- Whether the L4 service-source names (crafty-mc, unifi) should also go public. Not decided, not
+  changed.
+- `conformance/tests/l0-static/internal-services-publish-no-public-dns/check.sh` header lines 39–40
+  still describe the target as open. That is a sensitive path, so it is left for a follow-up.
@@ -0,0 +1,133 @@
+---
+date: 2026-10-04
+branch: feat/pihole-declare-state
+title: PiHole's operator-set state is now declared in git — the prerequisite for running more than one replica
+---
+
+## What changed
+
+Phase 1 of making LAN DNS survive a node failure. The operator asked for multi-replica PiHole;
+measurement showed that cannot be done safely first, because PiHole's important state was not
+reproducible from git. A second replica with a fresh PVC would have come up **serving DNS and
+blocking nothing, with a different admin password** — answers varying by which pod replied. This
+branch closes that; the replica work is phase 2 and is deliberately NOT here.
+
+**The undeclared-state inventory, measured not guessed.** `pihole.toml` marks every non-default
+setting `### CHANGED`, and those that came from the environment `### CHANGED (env)`. The complete
+live list on 2026-10-04:
+
+| Setting | Status |
+|---|---|
+| `dns.upstreams`, `dns.listeningMode`, `dns.rateLimit.count`, `misc.dnsmasq_lines` | `### CHANGED (env)` — already declared in `catalog/pihole.yaml` |
+| `files.macvendor` | `### CHANGED` — set by the image itself (container path), not operator state |
+| `webserver.api.pwhash` | **`### CHANGED` — set out of band. Declared NOWHERE.** |
+
+So exactly **one** operator-set value lived only on the PVC. Separately, the **adlists** were also
+undeclared: `gravity.db` held one list (StevenBlack's hosts) and 72,234 gravity domains, with nothing
+in this repo naming it.
+
+- **`apps/wave-0/pihole-bootstrap-secret.sops.yaml` (new, SOPS/age).** Secret `pihole-bootstrap`,
+  key `webui-password`. Named to the `<svc>-bootstrap-secret.sops.yaml` convention so
+  `render.py --emit-waves` places it with no renderer change. The value is a 32-char random password
+  generated at authoring time and **never printed to a transcript**.
+- **`catalog/pihole.yaml`:** `FTLCONF_webserver_api_password` sourced from that Secret.
+  `webserver.api.password` takes CLEARTEXT and FTL hashes it itself — verified in the running image,
+  where `pihole setpassword` *is* `setFTLConfigValue "webserver.api.password"`, commented upstream
+  "pihole-FTL will automatically hash the password". The PVC's `pwhash` becomes a derived artefact.
+- **`catalog/pihole.yaml`:** a `lifecycle.postStart` hook that **reconciles** the declared adlist set
+  into `gravity.db` on every container start — adds missing, re-enables disabled, and **disables**
+  (never deletes) anything enabled but undeclared, running `pihole -g` only when something moved.
+  Same level-triggered idiom, and the same reasoning, as `catalog/qbittorrent.yaml`'s postStart.
+- **`apps/wave-0/pihole.yaml`, `apps/wave-0/kustomization.yaml`:** re-rendered / secret registered.
+
+**Why a postStart hook and not a file.** PiHole v6 keeps adlists as DB rows. `/etc/pihole/adlists.list`
+still exists in `gravity.sh` but is **migration-only**: its import sits inside `migrate_to_database()`
+behind `if [ ! -e "$gravityDBfile" ]`, so it fires only while *creating* a database. A ConfigMap
+mounted there would be silently ignored on every pod this cluster runs today. There is also no
+`pihole adlist` CLI subcommand, so SQL is the actual interface; FTL ships an embedded shell
+(`pihole-FTL sqlite3`) so nothing had to be added to the image.
+
+**The password is kept rather than set empty, deliberately.** The admin UI is already behind the
+oauth2-proxy sidecar (`--allowed-group=platform-admins`), so an empty password would have been
+simpler. But the `pihole` Service still publishes **port 80** (PiHole direct, bypassing the proxy on
+`:4180`) and only the CiliumNetworkPolicy closes it — and this repo has a documented default-deny
+drift history. An unauthenticated admin UI that can rewrite LAN DNS should not rest on one policy
+staying correct. Narrowing that Service port is follow-up, not done here.
+
+**SENSITIVE PATHS — this PR adds a `**/*.sops.yaml`, so `ci/approval-policy.sh` will BLOCK it.** It
+needs the operator's `/approve-override <reason>` as a PR **review** of type "Comment". Not posted by
+me: CLAUDE.md section 1 forbids an agent clearing this on its own PR.
+
+## What I verified
+
+Live prod, 2026-10-04, read-only except where noted.
+
+- **The undeclared-state inventory above** — read from `pihole.toml`'s own `### CHANGED` markers, with
+  the password value masked. This is the measurement the whole branch rests on.
+- **A fresh `gravity.db` seeds NO adlists.** The image's own schema template
+  `/etc/.pihole/advanced/Templates/gravity.db.sql` creates the `adlist` table and contains no
+  `INSERT INTO adlist` at all — grepped in the running image. Confirmed empirically below (0 rows).
+  This is why the adlist declaration is load-bearing and not cosmetic.
+- **The adlists.list path really is migration-only** — read the guard in `/opt/pihole/gravity.sh`
+  (`migrate_to_database()`, `if [ ! -e "$gravityDBfile" ]`).
+- **`webserver.api.password` is a real settable key** — `pihole-FTL --config webserver.api.password`
+  returns a masked value, and `/usr/local/bin/pihole`'s `setpassword` sets exactly that key.
+- **THE RENDERED HOOK ACTUALLY WORKS — all four paths, extracted from `apps/wave-0/pihole.yaml` and
+  run in the live pod against THROWAWAY COPIES of `gravity.db` in `/tmp`** (production `gravity.db`
+  was never written, and `pihole -g` was stubbed so no 70k-domain download could start):
+  1. *Steady state* → `already matches the declared set - nothing to do`. No writes, no rebuild.
+  2. *True fresh DB built from the image's schema template* → 0 adlists before; hook added the
+     declared list and triggered the rebuild. **This is the fresh-PVC/new-replica case.**
+  3. *An undeclared list enabled* → disabled it (`enabled=0`), left the declared one alone, and did
+     not delete the row.
+  4. *Declared list present but disabled* → re-enabled it.
+- **Required tools exist in the image:** `bash`, `expr`, `grep`, `printf`, `sleep`, `pihole-FTL`.
+- **The Secret is genuinely encrypted** to the repo's age recipient; `stringData.webui-password`
+  starts `ENC[`. The gate's `no-plaintext-secrets (R-SEC-01)` L0 check passes.
+- **Schema:** `check-jsonschema` against `schemas/servicedefinition.schema.json` passes.
+- **Render determinism:** the canonical
+  `render.py --all catalog -o apps --wave-layout --no-image-automation` changed **only**
+  `apps/wave-0/pihole.yaml`. `render-golden` passes, so the hook's text also satisfies
+  `assert_substitution_safe` (no braced dollar expansion — the Flux envsubst contract).
+- **Toolbox gate:** rc **3**, with `spec-immutability`, `decisions-log`, `lint`, `l0`,
+  `render-golden`, `l1`, `approval-policy` all **pass**. The rc 3 is the two usual environmental
+  skips — `gitea-settings` (no local `SETTINGS_API_TOKEN`) and `conformance` L2/L3
+  (`CLUSTER_AVAILABLE` unset). **Not green; reported as rc 3.**
+
+## What I did NOT verify
+
+- **Nothing here has been reconciled.** Flux has not applied any of it, so the password has never
+  been set by this mechanism and the hook has never run during a real container start. Everything
+  above is the rendered artifact tested out-of-band, plus the image's own source.
+- **THE CURRENT ADMIN PASSWORD WILL STOP WORKING, and that is the intent.** FTL re-hashes
+  `webserver.api.password` on start, overwriting the existing `pwhash`. Whoever knows the present
+  password (nobody, which is why this started) loses it; the declared one replaces it. I have not
+  observed that replacement happen.
+- **Cold-start cost is unmeasured and could matter.** `postStart` must complete before the container
+  is reported Running. On a fresh PVC the hook waits for `gravity.db` (bounded, up to 180s) and then
+  runs a real `pihole -g` (72k domains, network downloads). `apps-wave-0` is `wait: true` and
+  `apps-wave-1` `dependsOn` it, so a slow cold start **delays the app tier**. The steady-state path
+  exits in about a second (scenario 1), so this is a rebuild/first-boot cost only — but it is not
+  measured, and the stub in testing means `pihole -g`'s real duration here is unknown.
+- **Whether `pihole -g` succeeds from inside the hook's environment.** Its failure is caught and the
+  existing `gravity.db` retained, so a failure degrades rather than breaks — untested live.
+- **Concurrency.** Nothing serialises the hook against FTL's own scheduled gravity update. SQLite
+  handles the locking, and the hook's writes are three small statements, but no interleaving was
+  exercised.
+- **MULTI-REPLICA — PHASE 2, NOT STARTED.** This branch only removes the blocker. The remaining work,
+  and it is substantial: a `replicas`/placement field on the ServiceDefinition schema (currently
+  `additionalProperties: false` with no such field), `render.py` emitting a StatefulSet with
+  `volumeClaimTemplates` (the config PVC is **ReadWriteOnce**, so a 2-replica Deployment cannot
+  work), pod anti-affinity, a PodDisruptionBudget (there is none today), re-rendered goldens, and
+  `conformance/**` updates — the last making phase 2 sensitive-path as well. The PVC rename a
+  StatefulSet forces will discard the existing `gravity.db`, which is only safe **because** of this
+  branch, and should be done only after this one is confirmed working in-cluster.
+- **The real risk the operator named is NOT fixed by phase 2 either.** Replicas address *node*
+  failure. LAN DNS still depends on the cluster as a whole — etcd quorum, Cilium, Longhorn, a bad
+  reconcile. Quantified on 2026-10-04: with `tolerationSeconds=300` plus node-NotReady detection
+  plus a Longhorn RWO force-detach, a failure of the node running the pod is roughly **6-10 minutes
+  of no LAN DNS**; a failure of the node holding the `pihole-dns-*` L2 lease (currently a
+  *different* node — lease on k3s-prod-2, pod on k3s-prod-1) is tens of seconds. An out-of-cluster
+  secondary resolver carrying the same `address=` lines is the only option that removes the
+  dependency, and it was considered and not chosen.
+- **Port 80 on the `pihole` Service is still published** and still closed only by policy. Not narrowed.
@@ -0,0 +1,117 @@
+---
+date: 2026-10-04
+branch: docs/pihole-ha-verified
+title: PiHole HA is VERIFIED on live prod — including a failover drill with zero dropped DNS queries
+---
+
+## What changed
+
+**No manifests changed. This entry exists to correct the log.** The three PiHole/Plex entries merged
+earlier today each say, correctly at the time, that nothing had been reconciled and that the outcome
+was unverified:
+
+- `decisions/2026-10-04-pihole-declare-state.md` — "Nothing here has been reconciled."
+- `decisions/2026-10-04-pihole-multi-replica.md` — "Nothing has been reconciled. No StatefulSet
+  exists, no second replica has ever run, the anti-affinity has never been evaluated..."
+
+All three are now **merged, reconciled and measured**, so those sentences are stale in exactly the way
+§2b of CLAUDE.md warns about — a reader who found them would conclude the work is unproven. This entry
+is the current state; it does not edit the earlier files (separate files, nothing to merge).
+
+Delivered and live as of 2026-10-04, from PRs #43 (plex tcp/32400), #44 (declare PiHole state) and
+#45 (PiHole 2 replicas + the mountPath bug):
+
+- PiHole is a **2-replica StatefulSet**, one pod per node, with per-replica Longhorn volumes, a
+  PodDisruptionBudget, and an `up` SLI that follows the rendered workload kind.
+- PiHole's admin password and adlists are **declared in git** and reproduce onto a fresh volume.
+- PiHole's config PVC is mounted where PiHole actually writes (`/etc/pihole`), which it was not before.
+
+## What I verified
+
+All measured against live prod on 2026-10-04, after the merge of #45 (`dce20b7`).
+
+**Reconcile.** `GitRepository` artifact `main@sha1:dce20b7`; `apps-wave-0`
+`lastAppliedRevision=main@sha1:dce20b7`, `Ready=True/ReconciliationSucceeded`. It completed in roughly
+**90 seconds**, well inside the Kustomization's `timeout: 5m` — the risk flagged in the multi-replica
+entry ("whether `pihole -g` finishes inside the 5 m timeout") did **not** materialise. A transient
+`Ready=False/DependencyNotReady: platform-config` seen mid-flight was the dependency cascade, not a
+failure; `platform-config` was already `Ready=True` at the new revision moments later.
+
+**Topology and objects.**
+
+| Measured | Value |
+|---|---|
+| pods / nodes | `pihole-0` → **k3s-prod-3**, `pihole-1` → **k3s-prod-2** (2/2 Running each) — the required anti-affinity was satisfied |
+| per-replica claims | `config-pihole-0`, `config-pihole-1` — both `Bound`, 2Gi |
+| old `pihole-config` PVC | **pruned** (it held only `lost+found`, so nothing was lost) |
+| old `pihole` Deployment | **pruned** — `No resources found in pihole namespace` |
+| PodDisruptionBudget | `minAvailable 1`, `disruptionsAllowed 1`, `currentHealthy 2` |
+
+**The mountPath bug is fixed.** `df` inside both pods reports `/etc/pihole` on
+`/dev/longhorn/pvc-...` with **6.9 M used** — previously it was on `overlay` while the Longhorn volume
+sat unused at `/data/config`. `backup.policy: daily-offcluster` now has real content to back up
+instead of reporting success over an empty volume.
+
+**Phase 1's declarations reproduce onto a FRESH volume — the point of doing it first.** Both replicas
+booted with an empty `/etc/pihole`:
+
+- **72,234 gravity domains** on each, rebuilt from the adlist declared in `catalog/pihole.yaml`.
+- **The declared SOPS secret authenticates on both pods.** `POST /api/auth` using the container's own
+  `FTLCONF_webserver_api_password` returned `"valid":true` on `pihole-0` **and** `pihole-1` — so
+  Secret → env → FTL's own hashing → working login, identically on both. (The secret's value was never
+  printed; an earlier reading of `pwhash` as "unset" was a quoting bug in the check, disproven by an
+  empty password being rejected with `password incorrect` and then by this positive test.)
+
+**User-facing behaviour through the VIP (192.168.2.23), from a LAN host.**
+
+    plex.home.keithjackson.dev   -> 192.168.2.20      split-horizon intact
+    home.keithjackson.dev        -> 192.168.2.20
+    unifi.home.keithjackson.dev  -> 192.168.2.21
+    google.com                   -> 142.250.115.102   recursion to 8.8.8.8 works
+    doubleclick.net              -> 0.0.0.0           blocking LIVE in FTL, not just in the DB
+    googleadservices.com         -> 0.0.0.0
+    12 consecutive queries       -> 12/12 answered
+
+**THE FAILOVER DRILL — observed by the operator, 2026-10-04.** A single StatefulSet replica was
+deleted. It was recreated in **~30 seconds**, and **DNS never dropped a request** throughout. This is
+the first direct evidence that the replica count delivers what it was added for, rather than merely
+rendering correctly.
+
+**The SLI is not inert, and re-deriving it mattered.** Measured in Prometheus:
+`kube_statefulset_status_replicas_ready{namespace="pihole", statefulset="pihole"}` returns **1 series,
+value 2**; `kube_deployment_status_replicas_available{namespace="pihole"}` now returns **0 series**;
+and Prometheus has `pihole-down` loaded as `state=inactive health=ok` with the StatefulSet expression.
+Had the rule stayed Deployment-shaped, its `or absent(...)` branch would have matched the now-absent
+series and **fired permanently** — a standing false page rather than silence. That is the concrete
+payoff of defaulting `livenessWorkload.kind` from the rendered kind.
+
+## What I did NOT verify
+
+- **NODE failure, as opposed to pod failure.** The drill deleted a pod, which does not exercise the
+  `tolerationSeconds: 300` path at all — that timer only applies when a node goes NotReady. The
+  mechanism protecting DNS is the same one just proven (the surviving replica plus
+  `externalTrafficPolicy: Cluster` on the DNS LoadBalancers), so user-visible impact should likewise be
+  nil; what the 300 s still delays is the *replacement* replica, and during that window there is no
+  redundancy left. Untested. Pulling power from a node is the only real test.
+- **DNS during the cutover itself.** I was not sampling resolution while #45 reconciled, so whether
+  there was a brief gap between the Deployment being pruned and the StatefulSet becoming Ready is
+  **unmeasured**. The drill above covers the steady-state failover case, not that one-off transition.
+- **Two-node loss.** The hard anti-affinity means `replicas: 2` needs two schedulable nodes; losing
+  two would leave a replica `Pending` and, with `wait: true`, stall the app tier. Reasoned, not
+  exercised — and it coincides with etcd quorum loss on a 3-node control plane.
+- **Longhorn capacity headroom.** Two 2Gi claims at 3 replicas each is ~12Gi where there was ~6Gi.
+  Both claims are `Bound` and healthy; free space was not checked.
+- **Query history is now split per replica** and the admin UI shows only whichever pod answered.
+  Adlists and password are identical, so blocking and auth are consistent; the *reporting* is not, and
+  whether that is acceptable long-term is the operator's call.
+- **The residual risk is unchanged, and it is the larger one.** Replicas address node failure. LAN DNS
+  still depends on the cluster — etcd quorum, Cilium, Longhorn, a bad reconcile — and only a resolver
+  outside the cluster carrying the same `address=` lines removes that dependency. Considered and not
+  chosen (operator decision, 2026-10-04).
+- **Plex remote access (#43) is NOT covered by this entry.** Its two manual steps (the WAN tcp/32400
+  forward and Plex's "manually specify public port") and the outside-in path remain as
+  `decisions/2026-10-03-plex-remote-access-l4.md` describes them. The operator reported it working;
+  I have not measured it from off-LAN.
+- **`--emit-waves` drift is still open.** Re-running it still wants to onboard `vm-demo.yaml` and
+  `keycloak.yaml`, both listed as intentionally not onboarded in `ci/render-golden.sh`. Unfixed; see
+  `decisions/2026-10-03-plex-remote-access-l4.md`.
@@ -0,0 +1,145 @@
+---
+date: 2026-10-04
+branch: feat/pihole-multi-replica
+title: PiHole runs two replicas on different nodes — and its config PVC was mounted where PiHole never writes
+---
+
+## What changed
+
+Phase 2 of making LAN DNS survive a node failure (phase 1:
+`decisions/2026-10-04-pihole-declare-state.md`). PiHole was one pod on one node, which for a
+whole-house resolver cost more than "until it reschedules" — measured 2026-10-04:
+
+| Failure | Recovery |
+|---|---|
+| the node running the pod dies | ~40s to mark NotReady + `tolerationSeconds: 300` before eviction + reschedule + a Longhorn RWO force-detach → **roughly 6-10 minutes with no LAN DNS** |
+| the node holding the `pihole-dns-*` L2 lease dies | Cilium re-elects an announcer, tens of seconds |
+
+Those were **different** nodes (lease on k3s-prod-2, pod on k3s-prod-1), so two of three nodes were
+single points of failure for LAN DNS with wildly different blast radii.
+
+**A new schema field, because this is intent and not a primitive.** `availability.replicas`
+(schema **1.10.0**, mechanism in `schemas/renderer-contract.md` `### availability`). `1` or absent
+renders a Deployment exactly as before — every pre-existing golden is byte-identical, verified. `>1`
+renders, together:
+
+- a **StatefulSet** instead of a Deployment. This is forced, not preferred: a longhorn PVC is
+  **ReadWriteOnce**, and one RWO claim cannot attach on two nodes, so a 2-replica Deployment does not
+  degrade — the second pod never schedules.
+- longhorn storage as **per-replica `volumeClaimTemplates`** (`config-pihole-0`, `config-pihole-1`);
+  the standalone PVC is no longer emitted, since it would be an orphan nothing mounts. A `bulk`
+  entry stays a standalone PVC — it is a shared static NFS export, not per-replica state.
+- **required** `podAntiAffinity` on `kubernetes.io/hostname`. `preferred` would let both replicas
+  land on one node and deliver none of the intended tolerance while reporting healthy.
+- `podManagementPolicy: Parallel`, so a cold start costs the max of the replicas' first-boot work
+  rather than the sum — `apps-wave-0` has `timeout: 5m`.
+- a **PodDisruptionBudget**, `minAvailable: 1`, so a node drain cannot take the last replica.
+- `observability.livenessWorkload.kind` defaulting to `StatefulSet`, so the `up` SLI reads
+  `kube_statefulset_status_replicas_ready`. A Deployment-shaped rule against a StatefulSet
+  `absent()`s forever — an alert that can never fire.
+
+`catalog/pihole.yaml` sets `replicas: 2`, not 3: two tolerates any single node loss and leaves a
+spare node for the evicted replica under the hard anti-affinity. Three would pin one per node with no
+spare, so a single node failure would leave a replica permanently `Pending` and — `apps-wave-0` being
+`wait: true` — stall the app tier for as long as the node was down.
+
+### The bug found on the way, and fixed here
+
+**PiHole's config PVC was mounted at a path PiHole never writes, so it held nothing and all of its
+state was ephemeral.** With no explicit `mountPath` the renderer defaults to `/data/<name>`; PiHole
+keeps everything in `/etc/pihole`. Measured on live prod 2026-10-04: `/etc/pihole` (7.1 M —
+`gravity.db`, `pihole.toml`, the query DB) was on the **overlay** filesystem, while the 2 Gi Longhorn
+volume sat at `/data/config` holding `lost+found` and 24 K of nothing. Consequences, all live until
+now:
+
+- every restart silently discarded PiHole's entire state and re-downloaded 72 k gravity domains;
+- `backup.policy: daily-offcluster` has been faithfully backing up an **empty** volume, which is
+  worse than no backup because it reports success;
+- it also means phase 1 was understated: that state was not merely undeclared, it was **ephemeral**.
+
+Same class of bug `catalog/plex.yaml` documents for its own `config` entry. Fixed in this branch
+rather than its own because with the wrong path the new per-replica volumes would be two 2 Gi
+Longhorn volumes (×3 Longhorn replicas each) holding only `lost+found` while all state stayed
+ephemeral — paying for persistence and getting none. It also removes this change's main operational
+risk: with `gravity.db` persisted, a restart is cheap instead of re-downloading 72 k domains inside a
+wave with a 5 m timeout.
+
+## What I verified
+
+- **Back-compatibility, the thing most likely to break quietly:** after the renderer change and
+  before `catalog/pihole.yaml` declared anything, the canonical
+  `render.py --all catalog -o apps --wave-layout --no-image-automation` produced **zero diff** across
+  the whole `apps/` tree. `render-golden` passes.
+- **The new behaviour is pinned by ~20 new `render.py --selftest` assertions** (so `ci/render-golden.sh`
+  hard-fails on a regression, and no `conformance/**` change was needed): `replicas: 1` still renders a
+  Deployment + standalone PVC + no PDB; `replicas: 2` renders a StatefulSet and **no** standalone
+  longhorn PVC, carries the replica count, omits `strategy`, sets `Parallel` and `serviceName`, names
+  the claim template exactly as the mounted volume, keeps RWO/longhorn on it, does **not** duplicate
+  that volume in `pod.volumes`, still mounts it on the container, renders required hostname
+  anti-affinity, renders exactly one `minAvailable: 1` PDB, and switches the SLI to
+  `kube_statefulset_status_replicas_ready`. A mixed longhorn+bulk case asserts the bulk share stays a
+  standalone PVC and an ordinary pod volume.
+- **Rendered pihole output, read back:** `replicas: 2`, `serviceName: pihole`,
+  `podManagementPolicy: Parallel`, no `strategy`, claim template `config` (RWO/longhorn/2Gi),
+  anti-affinity `kubernetes.io/hostname`, `pod.volumes: []`, mount `config -> /etc/pihole`,
+  PDB `minAvailable: 1`, SLI `kube_statefulset_status_replicas_ready`.
+- **The `/etc/pihole` mount is safe, evidenced from the image's own init rather than assumed:**
+  `/usr/bin/start.sh` is PID 1 and runs as **uid 0**; its line 5 is
+  `chown -R pihole:pihole /etc/pihole/ /var/log/pihole/`, so a fresh root-owned volume has its
+  ownership fixed before anything else starts (**no `fsGroup` needed, and one would be inert**); it
+  calls `migrate_gravity`, whose `migrate_to_database()` creates the gravity database when
+  `[ ! -e "$gravityDBfile" ]`; FTL writes `pihole.toml` itself from the `FTLCONF_*` env; and the
+  image's own default for `files.macvendor` was moved **out** of that directory, which is what a
+  maintainer does to a path they expect to be a volume.
+- **The PVC really is unused:** `df` showed `/etc/pihole` on `overlay` and
+  `/dev/longhorn/pvc-849edb86-...` on `/data/config`; `ls -A /data/config` → `lost+found` only.
+- **The gate caught a real regression and it is fixed.** The first run came back **rc 1**:
+  `alert-rules-unit-tested` failed because `apps/wave-0/tests/app-slis.test.yaml` fed pihole a
+  `kube_deployment_status_replicas_available` series and expected a `deployment: pihole` label. Three
+  input series, the firing expectation and the file header were updated;
+  `platform/monitoring/tests/run.sh apps/wave-0/tests/app-slis.test.yaml` → **SUCCESS** (6 rules).
+- **Schema:** `check-jsonschema` passes over all of `catalog/*.yaml`.
+- **Toolbox gate:** rc **3**, with `spec-immutability`, `decisions-log`, `lint`, `l0`,
+  `render-golden`, `l1`, `approval-policy` all **pass**. The rc 3 is the two environmental skips —
+  `gitea-settings` (no local `SETTINGS_API_TOKEN`) and `conformance` L2/L3 (`CLUSTER_AVAILABLE`
+  unset). **Not green; reported as rc 3.**
+- **No sensitive paths touched**, so unlike the previous two branches this one needs no
+  `/approve-override` — confirmed against `infra/gitea/approval-policy.json`.
+
+## What I did NOT verify
+
+- **Nothing has been reconciled.** No StatefulSet exists, no second replica has ever run, the
+  anti-affinity has never been evaluated by the scheduler, the PDB has never been consulted, and the
+  `/etc/pihole` mount has never been exercised on a real pod. Everything above is the rendered
+  artifact plus the image's own source.
+- **THE ROLLOUT WILL BRIEFLY DROP LAN DNS, and I have not measured for how long.** `postStart` must
+  complete before a container is reported Running, so a pod is not Ready — and the Service has no
+  endpoints — until phase 1's adlist hook finishes. On this transition both replicas start with an
+  empty `/etc/pihole`, so the hook finds a fresh `gravity.db` with zero adlists and runs a real
+  `pihole -g` (72 k domains, network download) before either pod becomes Ready. **Do this at a quiet
+  time.** Two mitigating facts: every pihole restart has *already* paid this cost (state was
+  ephemeral), and after this change `gravity.db` persists so later restarts should not. If it turns
+  out slow, backgrounding the gravity rebuild inside the hook would decouple readiness from it, at
+  the cost of a window where DNS resolves but does not block — not done here.
+- **Whether `pihole -g` finishes inside `apps-wave-0`'s 5 m `timeout`.** `Parallel` was chosen to
+  make this the max rather than the sum of two first boots, but the real duration is unmeasured (it
+  was stubbed during phase 1's testing). If the wave times out, the Kustomization reports failed even
+  though the pods may come up seconds later.
+- **The old PVC's deletion.** `apps-wave-0` has `prune: true`, so removing `pihole-config` from the
+  rendered output should delete it. That destroys nothing (it holds `lost+found`), but I have not
+  observed the prune, and I did not check whether Longhorn retains the backing volume afterwards.
+- **The transition overlap.** For a moment the pruned Deployment's pod and the new StatefulSet pods
+  may both match the Service selector. Both are configured identically from git so this should be
+  benign — unobserved.
+- **Longhorn capacity.** Two 2 Gi claims at 3 Longhorn replicas each is ~12 Gi of real storage where
+  there was ~6 Gi. Not checked against free space.
+- **Whether two independent query databases are acceptable to the operator.** Stats and query history
+  now split per replica, and the admin UI shows only the pod that answered. Adlists and password are
+  identical (phase 1), so blocking and auth are consistent; the *reporting* is not.
+- **The residual risk is unchanged and is the larger one.** Replicas address node failure. LAN DNS
+  still depends on the cluster — etcd quorum, Cilium, Longhorn, a bad reconcile — and only a resolver
+  outside the cluster carrying the same `address=` lines removes that. Considered and not chosen
+  (operator decision, 2026-10-04).
+- **`tolerationSeconds: 300` is untouched.** With two replicas the surviving one serves immediately, so
+  the 5-minute eviction delay no longer gates recovery — but it still delays the *replacement* replica,
+  and during that window there is no redundancy left.
@@ -0,0 +1,139 @@
+---
+date: 2026-10-04
+branch: fix/plex-advertise-plexdirect
+title: Plex vanished from the plex.tv account — every advertised connection was uncompletable; advertise the plex.direct name instead of the bare VIP
+---
+
+## What changed
+
+Reported symptom: the server no longer appears on the operator's plex.tv account; `plex.home.keithjackson.dev`
+shows "not authorized"; connecting straight to the VIP + port works.
+
+**Root cause, measured: `customConnections` held two hand-typed URLs and NEITHER could ever complete,
+so Plex had nothing publishable.** Because `customConnections` OVERRIDES Plex's own connection
+publishing, the server ended up advertising only dead endpoints. Live `Preferences.xml`, 2026-10-04:
+
+    customConnections="https://plex.home.keithjackson.dev:32400/,https://192.168.2.24:32400/"
+    ManualPortMappingMode="1"   ManualPortMappingPort=<ABSENT>
+
+- `https://192.168.2.24:32400` — **TLS can never validate.** Plex serves 32400 with
+  `CN=*.130072f20c334d6c8f63d0dfef87309b.plex.direct` and **no IP SAN**, so a bare-IP https URL fails
+  for every client. Not theoretical: Plex's log carries a continuous storm of
+  `CERT: incomplete TLS handshake from <node> ... sslv3 alert certificate unknown`, every few seconds,
+  from **real LAN clients** arriving through the new `plex-l4-tcp` VIP.
+- `https://plex.home.keithjackson.dev:32400` — **nothing listens.** The shared Gateway serves that host
+  on 443; measured `http=000` (connect failed) against `192.168.2.20:32400`.
+
+Consequence chain, each link measured: no completable connection -> plex.tv could publish nothing
+(`MyPlex: mapping state set to 'Mapped - Not Published'`, repeatedly) -> Plex fell back to **Plex Relay**
+(`startRelay`, remote forward to 127.0.0.1:32401) -> **the server disappeared from the account**, and
+clients reported "not authorized".
+
+The fix advertises Plex's own `<dashed-ip>.<serverhash>.plex.direct` form, which is public DNS resolving
+to the embedded address — so it points at the LAN VIP *and* matches the wildcard cert.
+
+- **`catalog/plex.yaml`:** `ADVERTISE_IP: ${PLEX_ADVERTISE_URL}` added, replacing the long-standing
+  DEFERRAL note. The deferral had weighed only the L7 Gateway URL; the operator chose the native L4 path
+  on 2026-10-03, so the open question was never "Gateway or not" but "which URL has a valid cert".
+- **`clusters/base/cluster-vars.yaml`:** new `PLEX_ADVERTISE_URL` knob (+ substitution-inventory comment
+  updated to name `plex: PLEX_ADVERTISE_URL`). The base default is an explicit **non-functional
+  placeholder** — dev is decommissioned and has no claimed server, and
+  `StrictPostBuildSubstitutions` makes an undefined var a hard failure, so the default only has to exist
+  and be correctly shaped.
+- **`clusters/prod/kustomization.yaml`:** prod value
+  `https://192-168-2-24.130072f20c334d6c8f63d0dfef87309b.plex.direct:32400`.
+- **`apps/wave-0/plex.yaml`:** re-rendered with the canonical command
+  (`python3 catalog/render.py --all catalog -o apps --wave-layout --no-image-automation`). Two added
+  lines, nothing else.
+
+**Why a whole-URL cluster-var and not something built from `${PLEX_L4_VIP}`:** postBuild substitution is
+plain envsubst and cannot turn `192.168.2.24` into `192-168-2-24`, and `<serverhash>` is the server's
+certificate identity, derivable from nothing in this tree. The dashed IP and `${PLEX_L4_VIP}` are two
+halves of one address and **nothing checks that they agree** — noted in both files.
+
+**SENSITIVE PATHS — this PR touches `clusters/**` (two files), so `ci/approval-policy.sh` will BLOCK it.**
+It needs the operator's `/approve-override <reason>` posted as a PR **review** of type "Comment".
+Deliberately not posted here: CLAUDE.md §1 forbids an agent clearing this on its own PR, and the agent
+account holds `owner`, so the prohibition is a rule rather than a limit.
+
+## What I verified
+
+All measurements against live prod, 2026-10-04, read-only. Secrets handled per CLAUDE.md §6: preference
+attributes were read by name-scoped grep, never by dumping `Preferences.xml` (it holds `PlexOnlineToken`),
+and the token was reported only as `<PRESENT,len=22>`.
+
+- **THE FIX ITSELF, with full TLS verification (no `curl -k`)** — the one measurement the whole change
+  rests on:
+  - `https://192-168-2-24.130072f20c334d6c8f63d0dfef87309b.plex.direct:32400/identity` -> **`http=200`, TLS verify OK**
+  - `https://192.168.2.24:32400/identity` -> **`http=000`, cert validation FAILED**
+  - that name resolves to `192.168.2.24` (`getent hosts`), confirming the dashed-IP mechanism works from
+    inside this network and is not blocked by DNS-rebinding protection on the resolver path tested.
+- **The server is still claimed — this was never a lost claim.** `/identity` -> `claimed="1"` with the real
+  `machineIdentifier`; `PlexOnlineToken` present, `PlexOnlineUsername="jakman85"`.
+- **The certificate, read off the live listener:** exactly one SAN,
+  `DNS:*.130072f20c334d6c8f63d0dfef87309b.plex.direct`. No IP SAN. This is why the bare-IP URL fails.
+- **The publish payload Plex actually sent to plex.tv**, which is what proves the connection list was the
+  problem rather than a guess: `PUT https://servers.plex.tv/devices/<id>?Connection[][uri]=https://plex.home.keithjackson.dev:32400&Connection[][uri]=https://192.168.2.24:32400&Connection[][uri]=http://136.51.115.252:32400&Connection[][uri]=http://10.0.2.136:32400&httpsEnabled=1...`
+  — all four unreachable from a LAN client or from plex.tv.
+- **`Mapped - Not Published`** at every reachability check across the whole log, and an active **Plex Relay**
+  session — the documented fallback when nothing is directly publishable.
+- **The `certificate unknown` TLS failures come from real client traffic, not a probe.** The source
+  `10.0.1.78` resolves via `cilium-dbg bpf ipcache get` to **identity 1 = `reserved:host`**, i.e. LAN
+  clients SNAT'd through `plex-l4-tcp` (`externalTrafficPolicy: Cluster`) on prod-3.
+- **The Gateway is NOT broken, and my first reading of it was wrong.** A 403 `Access denied` from
+  `server: envoy` on the Gateway path turned out to be an artifact of probing from inside the plex pod:
+  **all 16** Gateway hostnames return 403 from there, because Cilium's Gateway preserves the original
+  source identity and enforces the destination's ingress policy. Plex's own log shows a concurrent,
+  **working, signed-in** browser session through the Gateway returning 200s. The HTTPRoute is
+  `Accepted=True`/`ResolvedRefs=True` with no filters, and it is the only route claiming that hostname.
+- **Cilium is not denying the Gateway->Plex hop:** `cilium-dbg bpf policy get 3796` shows
+  `Allow Ingress reserved:ingress 32400/TCP`, PROXY PORT `NONE`, with 141k packets counted. No L7/HTTP
+  rules and no CCNPs exist anywhere in the cluster.
+- **In-cluster consumers were never affected:** Seerr's recently-added scans and watchlist syncs complete
+  cleanly with no auth errors; there are **zero** HTTP 401s in Plex's log (the five ` 401 ` hits are
+  transcoder segment numbers).
+- **`ADVERTISE_IP`'s mechanism, in the running image rather than from docs:**
+  `/etc/cont-init.d/40-plex-first-run` line 127 does `setPref "customConnections" "${ADVERTISE_IP}"`,
+  outside the claim guard, and `/.firstRunComplete` (lines 9/139) is on the container overlay, not the
+  `/config` PVC — so it re-applies on every pod start. Editing the env rolls the Deployment, which applies it.
+- **Toolbox gate** (`./infra/devcontainer/dev.sh run bash -c './ci/gate.sh'`): **rc 3**, with
+  `spec-immutability`, `decisions-log`, `lint`, `l0`, `render-golden`, `l1` and `approval-policy` all
+  **pass**. The rc 3 is the two known environmental skips — `gitea-settings` (no local
+  `SETTINGS_API_TOKEN`) and `conformance` L2/L3 (`CLUSTER_AVAILABLE` unset). **Not green. Reported as rc 3.**
+- **`render-golden` passing is the specific proof** that the catalog edit and the committed golden agree;
+  the rendered diff is exactly two lines and no other service's manifest moved.
+
+## What I did NOT verify
+
+- **THE OUTCOME. Nothing here has been observed to put the server back on the plex.tv account.** The
+  change is unreconciled: `ADVERTISE_IP` is absent from the live Deployment, the live
+  `customConnections` still holds the two broken URLs, and Plex has not been restarted. What is proven is
+  that the *new URL answers 200 with a valid certificate* and that *every currently advertised URL cannot
+  complete* — a necessary-and-plausibly-sufficient argument, not a demonstrated fix.
+- **Whether plex.tv will publish and list the server once this lands.** plex.tv's decision is made
+  off-network and cannot be measured from here. A LAN-only connection may still leave the account showing
+  the server as remotely unreachable even when local clients work.
+- **REMOTE (off-LAN) ACCESS IS NOT FIXED BY THIS, and may still be fully broken.** The advertised URL
+  resolves to an RFC1918 address, so it serves LAN clients only. The off-repo WAN tcp/32400 port-forward
+  (`runbooks/plex-remote-access.md`) is still **unverified** — `http://136.51.115.252:32400` was
+  unreachable from inside the LAN, but 2026-10-03 established there is no NAT hairpin here, so that is
+  **inconclusive**, not proof of absence. This repo has no off-LAN vantage point.
+- **`ManualPortMappingMode="1"` with `ManualPortMappingPort` ABSENT is left as-is.** Manual public-port
+  mapping is enabled with no port set — a half-configured state I did not change and whose exact effect on
+  publishing I did not isolate. There is **no GitOps path** to it (2026-10-03: the image supports no
+  `PLEX_PREFERENCE*` of any kind), so it is an operator UI step.
+- **Whether the operator intended the hand-typed URLs** to override the 2026-10-03 native-L4 design, or
+  typed them while debugging. I did not alter live preferences; the env var will overwrite
+  `customConnections` on the next pod start, which is a deliberate transfer of ownership from the web form
+  to git and will **silently discard** anything typed there since.
+- **The `plex.direct` hash's stability.** Treated as server identity and read once from the live cert. I
+  did not establish what reissues it, so this URL could go stale with no alert; nothing in CI or the
+  conformance suite checks that the dashed IP still matches `${PLEX_L4_VIP}` or that the hash is current.
+- **Resolver behaviour for the whole LAN.** `plex.direct` resolved correctly from a *pod* (cluster DNS
+  path). LAN clients resolve via PiHole, and dnsmasq's `stop-dns-rebind` would reject a public answer
+  carrying a private address. I did **not** test resolution from a LAN client, and this is the most likely
+  way the fix still fails.
+- **The CI approval-policy outcome.** The local stage reported green, but that is an offline N/A, not a
+  verdict: this PR touches `clusters/**`, so CI will block pending the operator's override.
+- **Why PMS restarted.** The process came up at 05:26 UTC while the pod is 2d2h old with 0 restarts, so
+  something restarted Plex inside the container. Unexplained and not investigated.
@@ -306,19 +306,27 @@ node_baseline_dns_servers:
  # k3s_base_domain, below). It reaches the issuer through the NODE resolver, i.e. through
  # these PUBLIC upstreams — not through CoreDNS's split-horizon answer.
  #
-  # THIS USED TO BE BROKEN, AND THE 2026-09-04 SUBDOMAIN SPLIT FIXED IT. While prod was
-  # authored to serve dev's hostnames, a prod node resolving keycloak.home.keithjackson.dev
-  # publicly got DEV's Gateway VIP (192.168.1.53) — the wrong cluster, on the wrong LAN, so
-  # kubectl OIDC SSO on prod could not work at all. Now prod's issuer is
-  # keycloak.prod.home.keithjackson.dev, and the public record for that name is published by
-  # PROD's own external-dns pointing at PROD's Gateway VIP (192.168.2.20). The public answer
-  # and the split-horizon answer therefore AGREE, and no /etc/hosts pin is needed.
+  # THE PUBLIC ANSWER IS NOW THE WAN, SO THE ISSUER IS PINNED (2026-10-03). Since
+  # platform/gateway/gateway.yaml set an external-dns `target`, every public name is a CNAME
+  # to wan.home.keithjackson.dev = the home WAN IP. The SOHO router does NOT hairpin: from the
+  # LAN, WAN:443 answers with the router's own self-signed placeholder cert (measured from
+  # k3s-prod-2/-3 2026-10-03: `curl` exit 60, ssl_verify_result=18). The apiserver would then
+  # be unable to fetch its issuer's discovery/JWKS, so kubectl OIDC SSO would fail on the
+  # next k3s restart or Keycloak key rotation. node_baseline_hosts_pins (below) pins the
+  # issuer host to the Gateway VIP in /etc/hosts, which glibc consults before these
+  # resolvers. (History: before the 2026-09-04 subdomain split the public answer was DEV's
+  # VIP; after it, the public answer was prod's own VIP and no pin was needed.)
  #
-  # STILL TO VERIFY on prod bring-up (PROD-PROVISIONING.md §10): this holds only once prod's
-  # external-dns has actually published the record. Until then the issuer does not resolve
-  # from the nodes and kubectl SSO fails — the cert-based admin kubeconfig is the break-glass
-  # path, and it always works. Ordering, not a design gap.
+  # The cert-based admin kubeconfig never depends on any of this. It is the break-glass path.
 node_baseline_dns_search: []
+# /etc/hosts pins: node-process names that must reach the in-cluster Gateway directly rather
+# than via the public (WAN, non-hairpinned) answer. See "THE PUBLIC ANSWER IS NOW THE WAN" above.
+# [parity] the IP MUST equal GATEWAY_VIP in clusters/prod/kustomization.yaml (192.168.2.20),
+# and the name MUST follow k3s_oidc_issuer_url's host (roles/k3s_server/defaults/main.yml).
+node_baseline_hosts_pins:
+  - ip: "192.168.2.20"
+    names:
+      - "keycloak.{{ k3s_base_domain }}"

 # --- platform hostname suffix, ANSIBLE SIDE (R-ENV-02, R-DNS-01) --------------------
 # PROD SERVES ITS OWN SUBDOMAIN. [parity] MUST equal BASE_DOMAIN in clusters/prod/
@@ -51,11 +51,14 @@ hubble:
 l2announcements:
  enabled: true

-# k3s ships its own cgroup mount; point Cilium at it instead of auto-mounting.
+# Point Cilium at the host's real cgroup2 root (/sys/fs/cgroup, mounted by systemd) instead of
+# auto-mounting. NOT /run/cilium/cgroupv2: nothing mounts cgroup2 there, so the agent mounted its
+# OWN namespaced cgroup and socket-LB covered only itself — host processes (the apiserver) lost
+# Service/LB translation. [parity] platform/cilium/helmrelease.yaml explains (2026-10-03).
 cgroup:
  autoMount:
    enabled: false
-  hostRoot: /run/cilium/cgroupv2
+  hostRoot: /sys/fs/cgroup

 # Roll the cilium agent DaemonSet ONE node at a time. [parity] must match the
 # platform/cilium HelmRelease. The cilium 1.20.1 chart defaults this to 2, which on a
@@ -242,3 +242,9 @@ node_baseline_dns_search:
 node_baseline_dns_options:
  - timeout:2
  - attempts:2
+# Static /etc/hosts entries (tags: dns, hosts), as a list of {ip, names}. EMPTY by default:
+# dev needs none. Prod pins its OIDC issuer host to the Gateway VIP because the public
+# answer for it is the non-hairpinned WAN IP (inventory/group_vars/k3s_prod/vars.yml).
+# Written as ONE marked block, so removing an entry here removes it from the node, and
+# an empty list removes the block entirely.
+node_baseline_hosts_pins: []
@@ -1090,3 +1090,22 @@
      {% endif %}
  when: node_baseline_dns_mechanism == 'resolv-conf-file'
  tags: [dns]
+
+# /etc/hosts pins — tag `dns` (and `hosts` on its own). Some node processes must reach a
+# public hostname at the in-cluster Gateway, not at its public answer. On prod, the
+# kube-apiserver's OIDC issuer resolves publicly to the WAN IP, which the router does not
+# hairpin. glibc reads /etc/hosts before any nameserver (nsswitch `files dns`), so a pin wins
+# regardless of which resolver mechanism above is in effect. One marked block: entries
+# dropped from node_baseline_hosts_pins disappear from the node, and an empty list
+# removes the block.
+- name: Pin node-process hostnames in /etc/hosts
+  ansible.builtin.blockinfile:
+    path: /etc/hosts
+    marker: "# {mark} node_baseline hosts pins (Ansible, tags: hosts)"
+    block: |
+      {% for pin in node_baseline_hosts_pins %}
+      {{ pin.ip }}	{{ pin.names | join(' ') }}
+      {% endfor %}
+    state: "{{ 'present' if node_baseline_hosts_pins | length > 0 else 'absent' }}"
+    backup: true
+  tags: [dns, hosts]
@@ -19,9 +19,9 @@ Serves **R-ENV-01** (you cannot certify a node you have not measured), **R-ENV-0

 ## Contract

- **Read-only and side-effect-free.** It reads `/sys`, `/proc`, `/etc`, runs `lsblk`/`lscpu`,
-  and — only when it can escalate — `smartctl`. It writes nothing to the node and
-  **installs nothing**. `smartmontools` is declared in `node_baseline_packages`; if it is
+- **Read-only and side-effect-free.** It reads `/sys`, `/proc`, `/etc`, runs `lsblk`/`lscpu`
+  (and `lspci` only if it happens to be installed), and — only when it can escalate —
+  `smartctl`. It writes nothing to the node and **installs nothing**. `smartmontools` is declared in `node_baseline_packages`; if it is
  absent the SMART checks SKIP with that as the stated reason (operator ruling 2026-08-31:
  package installation is `node_baseline`'s job, never a health check's, never a manual
  `apt-get`).
@@ -50,6 +50,9 @@ Serves **R-ENV-01** (you cannot certify a node you have not measured), **R-ENV-0
 | `memory.available`, `memory.swap`, `memory.ecc` | headroom, kubelet's swap-off prerequisite **and whether it survives a reboot** (see [Swap off is not the same as swap disabled](#swap-off-is-not-the-same-as-swap-disabled)), EDAC correctable/uncorrectable counters |
 | `memory.physical_total` | **is this machine using the RAM it has?** `/proc/zoneinfo` present-vs-managed pages (no root, metal or VM) and `MemTotal` vs the DIMM capacity SMBIOS reports (`dmidecode -t 17`, root, bare metal). Cause-independent by design — see [The EFI memattr check](#the-efi-memattr-check) |
 | `cpu.throttle`, `cpu.inventory`, `cpu.virtualisation` | thermal throttle events since boot, topology + frequency scaling, VT-x/AMD-V + `/dev/kvm` (KubeVirt, R-PLAT-03) |
+| `graphics.render_device` | is there a usable `/dev/dri/renderD*` character device — the path Plex hardware-transcodes through. See [The render-device check](#the-render-device-check) |
+| `graphics.drm_driver` | which DRM driver is bound to the display device, against `node_health_graphics_expected_driver` (`i915` here) |
+| `graphics.media_stack` | **INFO only**: `vcs*`/`vecs*` media engines, GuC/HuC firmware state, host VAAPI userspace. None of these is ever graded — see below for why |
 | `thermal.max` | hottest sensor across thermal zones and hwmon chips |
 | `pressure.cpu` / `.memory` / `.io` | PSI `some avg60` — dev learned to care about CPU PSI specifically |
 | `network.link`, `network.errors`, `network.resolvers` | link state/speed, NIC error counters, DNS upstream redundancy per address family |
@@ -91,6 +94,78 @@ used to blame that state on "this node has not been baselined yet", which sent t
 back to re-run the role that had already run.


+## The render-device check
+
+Plex transcodes on the iGPU through `/dev/dri/renderD128`. **When that node stops existing,
+nothing fails loudly.** Plex falls back to software transcoding by itself; the only symptom
+is a hot CPU and stuttering clients. A node rebuild, a BIOS change that disables the onboard
+IGD, a kernel that binds a different DRM driver, or a driver regression all produce exactly
+that silent degradation.
+
+The state was measured by hand on all three prod nodes on 2026-09-30 and found working. **A
+fact measured once decays silently**, so it is re-measured on every health run instead of
+living in a runbook. Nothing is installed, no module is loaded, no group is changed — the
+role's read-only contract is unchanged.
+
+| fact | where it comes from | root? |
+|---|---|---|
+| `/dev/dri` contents, ownership, mode, char-vs-not | `lstat()` on each entry | no |
+| owning group **name** for the render node's gid | `grp.getgrgid()` | no |
+| bound DRM driver | `/sys/class/drm/card*/device/driver` symlink target | no |
+| is the iGPU on the bus at all | `/sys/bus/pci/devices/*/class` (PCI base class `0x03`) | no |
+| media engines (`vcs*`, `vecs*`) | `/sys/class/drm/card*/engine/*` | no |
+| module load state | `/proc/modules` | no |
+| GuC/HuC | `/sys/module/i915/parameters/enable_guc` + `/proc/cmdline` | **parameter file is mode 0400** |
+
+### The PCI scan is separate from the DRM scan on purpose
+
+`/sys/class/drm/card*` exists only once a driver has **bound**. A machine whose iGPU sits on
+the bus with nothing attached has no `card0` at all — and that is precisely the failure being
+hunted. So the display device is looked for on the PCI bus independently, from sysfs, which
+is always there. That is what lets `graphics.render_device` distinguish
+
+- *no GPU in this machine* → **SKIP** (a VM guest; nothing to grade, the same convention
+  `thermal.max` uses for "no sensor on this hardware"), from
+- *a GPU is on the bus and there is still no render node* → **FAIL**.
+
+`lspci` is used **only** to decorate the PCI entry with a readable name, and it is treated as
+optional because it is: `pciutils` is **not** declared in `node_baseline_packages`, so nothing
+in this repo guarantees it. It did turn out to be present on all three prod nodes
+(`/usr/bin/lspci`, measured 2026-09-30 — Debian installs `pciutils` by default), but depending
+on that would mean depending on an undeclared package. Absent, `description` is `null`, the
+numeric IDs are printed instead, and no verdict changes.
+
+### The render node's GID is reported, never asserted
+
+The render device was `root:render`, **gid 992**, mode `0660` on all three prod nodes on
+2026-09-30. The check does **not** compare against `992`, and must not be changed to.
+`render` is a **systemd dynamic allocation** (`systemd-sysusers`), not a Debian static gid
+from `base-passwd`, so a rebuilt node can legitimately land on a different number. Plex's own
+container entrypoint stats the device and joins whichever group owns it, so drift genuinely
+is not a fault. Recording the observed group name and gid in the detail line makes drift
+**visible** without making it an error — a hard `!= 992` would be a false alarm waiting for a
+reinstall.
+
+What *is* configurable is the expected DRM driver, `node_health_graphics_expected_driver`
+(default `i915`), so the verdict stays out of task logic. `xe` bound instead of `i915` is a
+**WARN**, not a FAIL: the device still works, but the two drivers do not share a userspace
+media stack and a libva/iHD build that transcodes correctly against one is not automatically
+right for the other. On these nodes `xe.ko.xz` exists on disk and is bound to nothing.
+
+### Two things that are informational and must stay that way
+
+- **GuC/HuC firmware not loaded** (`i915.enable_guc=0`) is a known, accepted state here. It
+  costs low-power H.264 encode *quality* only; decode and encode both still run on the media
+  engines. Grading it would put a permanent WARN on a chosen configuration.
+- **No host VAAPI userspace / no `vainfo`** is **not a defect**. Plex ships its own libva and
+  media driver inside its container, so the host needs neither. Flagging it would invite
+  someone to `apt install vainfo` on a node to clear a health check — and this role installs
+  nothing.
+
+`enable_guc` is one of i915's `0400` module parameters, so an unprivileged pass usually
+cannot read it. It reports as an honest `NOT READ — <reason>` rather than a fabricated `0`,
+with `/proc/cmdline` as the world-readable second source for an explicit boot override.
+
 ## The EFI memattr check

 `infra/PROD-PROVISIONING.md` §0b is authoritative; this is only what the check needs.
@@ -273,3 +348,10 @@ stays `null` — no claim is made either way.
  but the wear/endurance context SKIPs rather than guessing.
 - `ioerr_cnt` counts commands the kernel saw fail, which includes benign rejected or
  unsupported commands during discovery. Attribution needs `dmesg`, which needs root.
+- `graphics.render_device`'s **FAIL path has never fired against a real node** — every prod
+  node has a working render device, and the role is read-only so one cannot be taken away to
+  prove it. The failure branches were exercised against synthetic fact sets only. The same
+  applies to the `xe`-bound WARN: no node here has `xe` bound.
+- `graphics.*` proves the render **device** exists and which driver owns it. It does **not**
+  prove a transcode succeeds — that needs a VAAPI round trip inside Plex's own container,
+  with Plex's own libva, which is out of scope for a node-level check that installs nothing.
@@ -35,6 +35,24 @@ node_health_try_privileged: true
 # --nocheck=standby: never spin up a sleeping disk just to read it. Read-only hygiene.
 node_health_smartctl_args: "--json=c --all --nocheck=standby"

+# --- Graphics / render device ---------------------------------------------------------
+# The DRM driver this site EXPECTS to be bound to its display device. Measured on all three
+# prod nodes 2026-09-30: `i915` on the CoffeeLake-S GT2 [UHD Graphics 630] iGPU of every HP
+# EliteDesk 800 G5 Mini. The newer `xe` module is present on disk as xe.ko.xz but is bound to
+# nothing, and that difference matters enough to surface: the two drivers do not share a
+# userspace media stack (a libva/iHD build that works against i915 is not automatically the
+# right one for xe), so a silent switch would change Plex's transcoding path. A site on AMD
+# or NVIDIA sets this to `amdgpu` / `nouveau` / `nvidia` rather than editing task logic.
+#
+# WHAT IS DELIBERATELY *NOT* A SETTING HERE: the numeric GID of the render node's owning
+# group. `render` is a systemd DYNAMIC allocation (systemd-sysusers), not a Debian static
+# one, so a rebuilt node may legitimately land on a different number — it was 992 on all
+# three prod nodes on 2026-09-30, and an assertion against that literal would be a false
+# alarm waiting to happen. Plex's own container entrypoint stats the device and joins
+# whichever group owns it, so drift genuinely is not a fault. The observed group name and GID
+# are therefore REPORTED in the check detail and never graded.
+node_health_graphics_expected_driver: i915
+
 # --- Thresholds ---------------------------------------------------------------------
 # All verdicts come from here, so a site can retune without touching task logic.
 node_health_thresholds:
@@ -35,10 +35,13 @@
 # =============================================================================

 import glob
+import grp
 import json
 import os
+import pwd
 import re
 import shutil
+import stat
 import subprocess
 import sys
 import time
@@ -756,6 +759,212 @@ def collect_pressure():
    return out


+# --- Graphics / DRM render device -----------------------------------------------------
+# WHY THIS IS HERE AT ALL. Plex transcodes on the iGPU through /dev/dri/renderD128. If that
+# node stops existing — a node rebuild, a BIOS change that disables the IGD, a kernel that
+# binds a different DRM driver — nothing breaks loudly: Plex silently falls back to software
+# transcoding and the only symptom is a hot CPU. So the render device is measured on every
+# health run rather than being a fact someone established once by hand.
+#
+# EVERYTHING BELOW IS UNPRIVILEGED. /dev/dri is world-traversable, the device nodes are
+# stat()able by anyone, /sys/class/drm and /sys/bus/pci/devices are world-readable, and so
+# are the module parameters under /sys/module. Nothing here needs root, so nothing here
+# belongs in the privileged pass. (GuC/HuC *load* status lives in debugfs, which does need
+# root AND debugfs mounted; the module PARAMETER below is the unprivileged stand-in and is
+# what actually decides whether the firmware is loaded at all.)
+#
+# THE PCI SCAN IS DELIBERATELY SEPARATE FROM THE DRM SCAN. /sys/class/drm/card* only exists
+# once a driver has BOUND. A machine whose iGPU is on the bus with no driver bound has no
+# card0 at all, and that is precisely the failure this is meant to catch — so the display
+# device is looked for on the PCI bus independently, from sysfs, which is always there.
+# `lspci` is used ONLY to decorate that with a human-readable name, and it is treated as
+# optional because it IS optional: pciutils is not declared in node_baseline_packages, so
+# nothing in this repo guarantees it. It did turn out to be present on all three prod nodes
+# (/usr/bin/lspci, measured 2026-09-30 — Debian installs pciutils by default), but relying on
+# that would make the collector depend on an undeclared package. A missing lspci leaves
+# `description` null, the numeric IDs are reported instead, and no verdict changes.
+DISPLAY_PCI_CLASS_PREFIX = "0x03"  # PCI base class 03 = display controller (VGA/3D/other)
+# DRM drivers worth reporting the load state of. The bound-driver symlink is authoritative
+# (a built-in driver never appears in /proc/modules); this is corroboration only, and it is
+# how "xe.ko.xz exists on disk but is bound to nothing" becomes a visible fact.
+DRM_MODULES = ("i915", "xe", "amdgpu", "radeon", "nouveau", "nvidia")
+
+
+def _uid_name(uid):
+    try:
+        return pwd.getpwuid(uid).pw_name
+    except (KeyError, OverflowError, TypeError, OSError):
+        return None
+
+
+def _gid_name(gid):
+    """Group NAME for a numeric gid, or None when nothing resolves it.
+
+    Reported as an OBSERVED value, never asserted against a literal. The `render` group is
+    a systemd DYNAMIC allocation (it is created by systemd-sysusers, not fixed by Debian's
+    static base-passwd list), so its gid can legitimately differ after a rebuild. Plex's own
+    container init stats the device and joins whatever group owns it, so a moved gid is
+    drift to notice, not a fault to fail on."""
+    try:
+        return grp.getgrgid(gid).gr_name
+    except (KeyError, OverflowError, TypeError, OSError):
+        return None
+
+
+def _module_param(mod, param):
+    """One /sys/module/<mod>/parameters/<param>, with WHY it is missing when it is.
+
+    Three outcomes that must not be conflated: a real value; "the module exposes no such
+    parameter" (not loaded, built in, or a driver that simply has no such knob — xe has no
+    enable_guc, GuC is mandatory there); and "the file exists but is mode 0400". Most i915
+    module parameters are `i915_param_named(..., 0400, ...)` — root-only — so an
+    unprivileged pass legitimately cannot read them. Returning null with the reason keeps
+    that an honest UNKNOWN instead of a fabricated 0. read_text() also records the
+    permission denial in UNREADABLE, so the summary lists it with everything else root
+    would add."""
+    path = "/sys/module/%s/parameters/%s" % (mod, param)
+    if not os.path.exists(path):
+        return {"value": None, "reason": "%s exposes no %s parameter (module not loaded, "
+                                         "built in, or no such knob on this driver)" % (mod, param)}
+    value = read_text(path)
+    if value is None:
+        return {"value": None, "reason": "%s is not readable without root (i915 module "
+                                         "parameters are mode 0400)" % path}
+    return {"value": value, "reason": None}
+
+
+def _dev_node(path):
+    """stat() one /dev/dri entry. lstat, so a symlink is reported as the symlink it is."""
+    try:
+        st = os.lstat(path)
+    except OSError as exc:
+        UNREADABLE.append({"path": path, "reason": "stat failed: %s" % exc})
+        return None
+    return {
+        "path": path,
+        "name": os.path.basename(path),
+        "char_device": stat.S_ISCHR(st.st_mode),
+        "mode": "%04o" % stat.S_IMODE(st.st_mode),
+        "uid": st.st_uid,
+        "gid": st.st_gid,
+        "owner": _uid_name(st.st_uid),
+        "group": _gid_name(st.st_gid),
+    }
+
+
+def collect_graphics():
+    dri_dir = "/dev/dri"
+    dri_present = os.path.isdir(dri_dir)
+    devices = []
+    if dri_present:
+        for path in sorted(glob.glob(dri_dir + "/*")):
+            # /dev/dri/by-path/ is a directory of symlinks, not a device. Skipping
+            # directories rather than filtering to char devices is deliberate: anything
+            # in here that is NOT a character special file is an anomaly worth seeing, and
+            # `char_device: false` is how it becomes visible.
+            if os.path.isdir(path):
+                continue
+            entry = _dev_node(path)
+            if entry is not None:
+                devices.append(entry)
+    render_nodes = [d for d in devices if d["name"].startswith("renderD")]
+
+    # Bound DRM devices. /sys/class/drm also holds one entry per CONNECTOR
+    # (card0-HDMI-A-1, card0-DP-1 ...) — those are outputs, not devices, and are skipped.
+    cards = []
+    for cdir in sorted(glob.glob("/sys/class/drm/card[0-9]*")):
+        name = os.path.basename(cdir)
+        if "-" in name:
+            continue
+        driver = None
+        drvlink = cdir + "/device/driver"
+        if os.path.islink(drvlink):
+            driver = os.path.basename(os.path.realpath(drvlink))
+        address = None
+        devlink = cdir + "/device"
+        if os.path.islink(devlink):
+            address = os.path.basename(os.path.realpath(devlink))
+        # i915/xe publish one directory per hardware engine. vcs* is a video (H.264/HEVC)
+        # engine and vecs* the video-enhancement engine: those two are what a transcode
+        # actually runs on, so their presence is the difference between "a GPU is bound"
+        # and "this GPU can do media work".
+        engines = sorted(
+            {
+                os.path.basename(p)
+                for p in (glob.glob(cdir + "/engine/*") + glob.glob(cdir + "/device/engine/*"))
+                if os.path.isdir(p)
+            }
+        )
+        cards.append(
+            {
+                "card": name,
+                "driver": driver,
+                "pci_address": address,
+                "vendor_id": read_text(cdir + "/device/vendor"),
+                "device_id": read_text(cdir + "/device/device"),
+                "engines": engines,
+                "media_engines": [e for e in engines if e.startswith(("vcs", "vecs"))],
+            }
+        )
+
+    # Display-class devices on the PCI bus, driver bound or not.
+    pci_display = []
+    for pdir in sorted(glob.glob("/sys/bus/pci/devices/*")):
+        class_id = read_text(pdir + "/class") or ""
+        if not class_id.startswith(DISPLAY_PCI_CLASS_PREFIX):
+            continue
+        address = os.path.basename(pdir)
+        driver = None
+        if os.path.islink(pdir + "/driver"):
+            driver = os.path.basename(os.path.realpath(pdir + "/driver"))
+        description = None
+        rc, out, _ = run(["lspci", "-nn", "-s", address])
+        if rc == 0 and out.strip():
+            description = out.strip().splitlines()[0]
+        pci_display.append(
+            {
+                "address": address,
+                "class_id": class_id or None,
+                "vendor_id": read_text(pdir + "/vendor"),
+                "device_id": read_text(pdir + "/device"),
+                "revision": read_text(pdir + "/revision"),
+                "driver": driver,
+                "description": description,
+            }
+        )
+
+    loaded = set()
+    for line in (read_text("/proc/modules") or "").splitlines():
+        fields = line.split()
+        if fields:
+            loaded.add(fields[0])
+    modules = {m: (m in loaded) for m in DRM_MODULES}
+
+    # GuC/HuC. i915's `enable_guc` is a bitmask: 0 = neither, 1 = GuC submission, 2 = HuC
+    # firmware, 3 = both, -1 = whatever the driver defaults to on this platform. It is the
+    # one place the state is visible without root AND without debugfs — but the sysfs file
+    # is mode 0400, so unprivileged it usually reads back as an honest UNKNOWN. The kernel
+    # command line is the world-readable second source: an explicit `i915.enable_guc=` there
+    # is authoritative about what was ASKED for, even when the parameter file cannot be read.
+    uc = {mod: {"enable_guc": _module_param(mod, "enable_guc")} for mod in ("i915", "xe")}
+    cmdline_params = [
+        token
+        for token in (read_text("/proc/cmdline") or "").split()
+        if token.startswith(tuple("%s." % m for m in DRM_MODULES))
+    ]
+
+    return {
+        "dri_present": dri_present,
+        "devices": devices,
+        "render_nodes": render_nodes,
+        "drm_cards": cards,
+        "pci_display": pci_display,
+        "modules": modules,
+        "uc": uc,
+        "cmdline_params": cmdline_params,
+    }
+
+
 # --- Storage ------------------------------------------------------------------------
 PSEUDO_FS = {
    "proc", "sysfs", "devtmpfs", "devpts", "tmpfs", "cgroup", "cgroup2", "pstore",
@@ -1095,7 +1304,15 @@ def collect_os():
 # we only DETECT, and the role skips the SMART checks with an accurate reason.
 def collect_tools():
    tools = {}
-    for tool in ("smartctl", "nvme", "ethtool", "dmidecode", "lsblk", "lscpu"):
+    # lspci (pciutils) and vainfo (libva-utils) are BOTH optional and NEITHER is declared in
+    # node_baseline_packages, so nothing in this repo guarantees either. Measured on prod
+    # 2026-09-30: lspci IS there (/usr/bin/lspci — Debian ships pciutils by default) and
+    # vainfo is NOT. lspci only decorates the PCI display entry with a readable name; vainfo
+    # is not needed on the host at all, because Plex ships its own libva and media driver
+    # inside its container. They are probed so the report can state their presence or absence
+    # as a fact instead of leaving the reader to wonder — and so nothing here is ever tempted
+    # to install them. This role installs nothing.
+    for tool in ("smartctl", "nvme", "ethtool", "dmidecode", "lsblk", "lscpu", "lspci", "vainfo"):
        tools[tool] = which(tool)
    return tools

@@ -1127,6 +1344,7 @@ def main():
        "os": collect_os(),
        "dmi": collect_dmi(),
        "cpu": collect_cpu(),
+        "graphics": collect_graphics(),
        "memory": collect_memory(),
        "thermal": collect_thermal(),
        "pressure": collect_pressure(),
@@ -652,6 +652,204 @@
         ~ (', lscpu Virtualization=' ~ c.virtualization if c.virtualization else '')
         ~ (', running as a guest of ' ~ c.hypervisor_vendor if c.is_virtual_guest | bool else '') }}

+# --- Graphics / render device ----------------------------------------------------------
+# WHY A HEALTH CHECK AND NOT A NOTE IN A RUNBOOK. Plex transcodes on the iGPU through
+# /dev/dri/renderD128. When that node stops existing nothing fails loudly — Plex falls back
+# to software transcoding on its own and the only symptom is a hot CPU and stuttering
+# clients. A node rebuild, a BIOS setting that disables the onboard IGD, or a kernel that
+# binds a different DRM driver all produce exactly that silent degradation. It was measured
+# by hand on all three prod nodes on 2026-09-30 and found working; a fact measured once
+# decays silently, so it is re-measured on every health run instead.
+#
+# READ-ONLY, AND UNPRIVILEGED. Every fact behind these three checks comes from the
+# unprivileged pass: stat() on /dev/dri, the /sys/class/drm bound-driver symlink,
+# /sys/bus/pci/devices class codes, /proc/modules and /proc/cmdline. The role installs
+# nothing (no vainfo, no pciutils), loads no module and changes no group — see the role
+# README. The ONE fact that is root-gated is i915's enable_guc module parameter (mode 0400),
+# and it reports as an honest UNKNOWN with the reason rather than as a 0.
+#
+# THE THREE VERDICTS ARE SPLIT ON PURPOSE:
+#   graphics.render_device  the OUTCOME — can anything hardware-transcode on this node.
+#   graphics.drm_driver     the CAUSE side — what is bound, and is it what this site expects.
+#   graphics.media_stack    INFO only. GuC/HuC and host VAAPI userspace both belong here and
+#                           NEITHER is ever a failure; see the task's own comment for why.
+- name: Assess — locate the display device and its bound driver (once, shared by three checks)
+  ansible.builtin.set_fact:
+    node_health_render:
+      # Display-class PCI devices (base class 0x03), driver bound or not. This is the
+      # question "is there a GPU in this machine at all", and it is answered from the PCI
+      # bus rather than from /sys/class/drm because a card* entry only exists once a driver
+      # has BOUND — so a GPU with nothing bound to it has no card* and would otherwise look
+      # identical to a machine with no GPU.
+      display_devices: "{{ g.pci_display }}"
+      bound_drivers: >-
+        {{ g.pci_display | map(attribute='driver') | select('string') | unique | list }}
+      # Only a CHARACTER SPECIAL file is a usable render node. Anything else called renderD*
+      # is an anomaly, not a device, and must not be counted as one.
+      render_chardevs: "{{ g.render_nodes | selectattr('char_device') | list }}"
+      # Video (vcs*) and video-enhancement (vecs*) engines across every bound card — the
+      # difference between "a GPU is bound" and "this GPU can do media work".
+      media_engines: >-
+        {{ g.drm_cards | map(attribute='media_engines') | flatten | unique | sort | list }}
+  vars:
+    ansible_become: false
+    g: "{{ node_health_facts.graphics }}"
+
+- name: Assess — DRM render node for hardware transcoding
+  ansible.builtin.set_fact:
+    node_health_checks: "{{ node_health_checks + [{
+      'id': 'graphics.render_device',
+      'title': 'DRM render node present for hardware transcoding (/dev/dri/renderD*)',
+      'status': render_status,
+      'detail': render_detail }] }}"
+  vars:
+    ansible_become: false
+    g: "{{ node_health_facts.graphics }}"
+    r: "{{ node_health_render }}"
+    # SKIP, not PASS and not FAIL, when the machine has no display device at all: there is
+    # then nothing to grade, and the same convention thermal.max already uses for "no sensor
+    # on this hardware" applies. FAIL is reserved for the state that actually breaks Plex —
+    # a GPU IS on the bus and there is still no render node.
+    render_status: >-
+      {{ 'SKIP' if (r.display_devices | length) == 0
+         else ('PASS' if (r.render_chardevs | length) > 0 else 'FAIL') }}
+    dri_listing: >-
+      {{ (g.devices | map(attribute='name') | join(', ')) if (g.devices | length) > 0
+         else 'nothing' }}
+    render_detail: >-
+      {%- if (r.display_devices | length) == 0 -%}
+      this machine exposes no display-class PCI device (base class 0x03) at all, so there is
+      no GPU here for a render node to belong to and hardware transcoding is not available on
+      this node by construction. Normal for a VM guest. Measured from /sys/bus/pci/devices,
+      not assumed.
+      {%- elif (r.render_chardevs | length) > 0 -%}
+      {{ r.render_chardevs | length }} render node(s):
+      {% for d in r.render_chardevs %}{{ d.path }} (character special, mode {{ d.mode }},
+      owner {{ d.owner | default(d.uid, true) }}, group {{ d.group | default('unresolved', true) }}
+      gid {{ d.gid }}){{ ', ' if not loop.last else '' }}{% endfor %}.
+      /dev/dri holds: {{ dri_listing }}.
+      THE OWNING GROUP AND GID ARE REPORTED AS OBSERVED, NEVER ASSERTED against a literal:
+      `render` is a systemd DYNAMIC allocation rather than a Debian static one, so a rebuilt
+      node may legitimately land on a different number (all three prod nodes read gid 992 on
+      2026-09-30), and Plex's container entrypoint stats the device and joins whichever group
+      owns it. Drift is therefore visible in this line without being a failure.
+      {%- else -%}
+      HARDWARE TRANSCODING IS BROKEN ON THIS NODE. A display-class device IS present on the
+      PCI bus ({{ r.display_devices | map(attribute='address') | join(', ') }}) but there is
+      no usable character-special render node: /dev/dri
+      {{ 'does not exist' if not (g.dri_present | bool) else 'holds ' ~ dri_listing }}.
+      {{ ('Note that /dev/dri holds ' ~ (g.render_nodes | map(attribute='name') | join(', '))
+      ~ ' — present, but NOT character special, which is an anomaly in its own right. ')
+         if (g.render_nodes | length) > 0 else '' }}
+      Nothing fails loudly in this state — Plex silently falls back to software transcoding
+      and the only symptom is CPU load. Check graphics.drm_driver below for whether a DRM
+      driver bound at all, then whether the onboard IGD is still enabled in BIOS.
+      {%- endif -%}
+
+- name: Assess — DRM driver bound to the display device
+  ansible.builtin.set_fact:
+    node_health_checks: "{{ node_health_checks + [{
+      'id': 'graphics.drm_driver',
+      'title': 'DRM driver bound to the display device (expected ' ~ expected ~ ')',
+      'status': driver_status,
+      'detail': driver_detail }] }}"
+  vars:
+    ansible_become: false
+    g: "{{ node_health_facts.graphics }}"
+    r: "{{ node_health_render }}"
+    expected: "{{ node_health_graphics_expected_driver }}"
+    # WARN rather than FAIL when a DIFFERENT driver is bound (the `xe` case): the device
+    # still works, but the userspace media stack is not the same one, so it is a change the
+    # operator must see before a transcode regression surfaces as a user complaint. FAIL is
+    # for nothing bound at all, which is the cause of a missing render node.
+    driver_status: >-
+      {{ 'SKIP' if (r.display_devices | length) == 0
+         else ('FAIL' if (r.bound_drivers | length) == 0
+               else ('PASS' if expected in r.bound_drivers else 'WARN')) }}
+    driver_detail: >-
+      {%- if (r.display_devices | length) == 0 -%}
+      no display-class PCI device on this machine, so no DRM driver can be bound to one.
+      Measured from /sys/bus/pci/devices.
+      {%- else -%}
+      {% for p in r.display_devices %}{{ p.address }}
+      {{ p.description | default(((p.vendor_id | default('?', true)) ~ ':' ~ (p.device_id | default('?', true))), true) }}
+      -> driver {{ p.driver | default('NONE BOUND', true) }}{{ '; ' if not loop.last else '' }}{% endfor %}.
+      {{ ('DRM cards: ' ~ (g.drm_cards | map(attribute='card') | zip(g.drm_cards | map(attribute='driver'))
+           | map('join', '=') | join(', ')) ~ '. ') if (g.drm_cards | length) > 0
+         else 'No /sys/class/drm/card* entry exists, which is consistent with no driver being bound. ' }}
+      Module load state (/proc/modules, corroboration only — a built-in driver never appears
+      there, the bound-driver symlink above is what decides):
+      {{ g.modules | dict2items | map(attribute='key') | zip(g.modules | dict2items | map(attribute='value') | map('string'))
+         | map('join', '=') | join(', ') }}.
+      {%- if (r.bound_drivers | length) == 0 %}
+      NOTHING IS BOUND to the display device. That is why there is no render node: the DRM
+      subsystem creates /dev/dri/renderD* only when a driver attaches. Check whether the
+      module exists for this kernel and whether it was blacklisted or failed to probe.
+      {%- elif expected in r.bound_drivers %}
+      This is the expected driver for this site (node_health_graphics_expected_driver).
+      {%- else %}
+      EXPECTED `{{ expected }}`, GOT `{{ r.bound_drivers | join(', ') }}`. The device still
+      works, so this is not an outage — but the two DRM drivers do not share a userspace
+      media stack, and a libva/iHD build that transcodes correctly against one is not
+      automatically right for the other. Confirm Plex still hardware-transcodes before
+      treating this as benign.
+      {%- endif %}
+      {%- endif -%}
+
+# INFO ONLY, AND THAT IS A DELIBERATE RULING, NOT AN OVERSIGHT:
+#
+#   GuC/HuC firmware not loaded (i915.enable_guc=0) — the accepted state on these nodes. It
+#   costs low-power H.264 encode QUALITY, nothing else; decode and encode both still work
+#   through the media engines. Grading it would put a permanent WARN on a configuration the
+#   operator has chosen.
+#   No host VAAPI userspace / no `vainfo` — NOT a defect. Plex ships its own libva and media
+#   driver inside its container, so the host needs neither. Flagging it would invite someone
+#   to `apt install vainfo` on a node to clear a health check, and this role installs nothing.
+#
+# The media ENGINES are the fact worth watching in this line: vcs* (video) and vecs*
+# (video-enhancement) are what a transcode actually runs on.
+- name: Assess — media engines, GuC/HuC firmware and host VAAPI userspace (informational)
+  ansible.builtin.set_fact:
+    node_health_checks: "{{ node_health_checks + [{
+      'id': 'graphics.media_stack',
+      'title': 'Media engines, GuC/HuC firmware and host VAAPI userspace (informational)',
+      'status': ('SKIP' if (node_health_render.display_devices | length) == 0 else 'INFO'),
+      'detail': media_detail }] }}"
+  vars:
+    ansible_become: false
+    g: "{{ node_health_facts.graphics }}"
+    r: "{{ node_health_render }}"
+    # `.get()` rather than a subscript: uc only carries the two Intel drivers, and a site
+    # that set the expected driver to amdgpu must get a clean "no such parameters" line
+    # instead of an undefined-key traceback.
+    guc: "{{ g.uc.get(node_health_graphics_expected_driver) }}"
+    guc_text: >-
+      {{ ('enable_guc=' ~ guc.enable_guc.value) if (guc is mapping and guc.enable_guc.value is not none)
+         else ('NOT READ — ' ~ (guc.enable_guc.reason | default('no reason recorded', true))
+               if guc is mapping
+               else 'no /sys/module/' ~ node_health_graphics_expected_driver ~ '/parameters at all') }}
+    media_detail: >-
+      {%- if (r.display_devices | length) == 0 -%}
+      no display-class PCI device on this machine, so there is no media engine, no GuC/HuC
+      firmware and nothing to transcode with.
+      {%- else -%}
+      media engines:
+      {{ (r.media_engines | join(', ')) if (r.media_engines | length) > 0
+         else 'NONE exposed under /sys/class/drm/card*/engine — a bound driver with no vcs*/vecs* engine cannot do media work' }}
+      (all engines: {{ (g.drm_cards | map(attribute='engines') | flatten | unique | sort | join(', ')) | default('none', true) }}).
+      GuC/HuC: {{ guc_text }}.
+      {{ ('Kernel cmdline carries ' ~ (g.cmdline_params | join(' '))) if (g.cmdline_params | length) > 0
+         else 'No i915/xe parameter on the kernel cmdline' }}.
+      GuC/HuC being off is a KNOWN, ACCEPTED state here and is never graded: it costs
+      low-power H.264 encode quality only, and both decode and encode still run on the
+      engines above.
+      Host VAAPI userspace: vainfo
+      {{ ('present at ' ~ node_health_facts.tools.vainfo) if node_health_facts.tools.vainfo
+         else 'NOT installed on this host' }}
+      — also never graded, because Plex ships its own libva and media driver inside its
+      container, so the host needs none. This role installs nothing.
+      {%- endif -%}
+
 # --- Thermals -----------------------------------------------------------------------
 - name: Assess — temperatures
  ansible.builtin.set_fact:
@@ -136,6 +136,37 @@
 {% endfor %}
 {% endfor %}

+--------------------------------------------------------------------------------
+ GRAPHICS / RENDER DEVICE (the Plex hardware-transcoding path)
+--------------------------------------------------------------------------------
+{% set gfx = node_health_facts.graphics %}
+{% if gfx.pci_display %}
+{% for p in gfx.pci_display %}
+ pci  {{ '%-12s' | format(p.address) }} {{ p.description | default(((p.vendor_id | default('?', true)) ~ ':' ~ (p.device_id | default('?', true)) ~ ' class ' ~ (p.class_id | default('?', true))), true) }}
+              driver {{ p.driver | default('NONE BOUND', true) }}
+{% endfor %}
+{% else %}
+ pci         no display-class device (PCI base class 0x03) on this machine — no GPU here
+{% endif %}
+{% for c in gfx.drm_cards %}
+ drm  {{ '%-12s' | format(c.card) }} driver {{ '%-8s' | format(c.driver | default('-', true)) }} at {{ c.pci_address | default('-', true) }}  engines {{ c.engines | join(' ') | default('none exposed', true) }}
+{% endfor %}
+{% if gfx.dri_present %}
+{% for d in gfx.devices %}
+ dev  {{ '%-12s' | format(d.name) }} {{ '%-18s' | format('character special' if d.char_device else 'NOT A CHAR DEVICE') }} mode {{ d.mode }}  {{ d.owner | default(d.uid, true) }}:{{ d.group | default('(gid unresolved)', true) }} gid {{ d.gid }}
+{% endfor %}
+{% else %}
+ dev         /dev/dri DOES NOT EXIST — no DRM device node on this node at all
+{% endif %}
+{# GID is printed, never asserted: `render` is a systemd dynamic allocation, so a rebuild may
+   legitimately move it and Plex's container init joins whichever group owns the device. See
+   defaults/main.yml (node_health_graphics_expected_driver) for the full reasoning. #}
+{% set guc = gfx.uc.get(node_health_graphics_expected_driver) %}
+ GuC/HuC     {{ ('enable_guc=' ~ guc.enable_guc.value) if (guc and guc.enable_guc.value is not none) else (('NOT READ — ' ~ (guc.enable_guc.reason | default('no reason recorded', true))) if guc else ('no /sys/module/' ~ node_health_graphics_expected_driver ~ '/parameters')) | regex_replace('\s+', ' ') | trim | wordwrap(64) | indent(14) }}
+             {{ ('kernel cmdline: ' ~ (gfx.cmdline_params | join(' '))) if gfx.cmdline_params else 'no i915/xe parameter on the kernel cmdline' }}
+ modules     {% for m, loaded in gfx.modules.items() %}{{ m }}={{ loaded | string | lower }}{{ ', ' if not loop.last }}{% endfor %} (/proc/modules — corroboration only; the bound-driver symlink above decides)
+ vainfo      {{ node_health_facts.tools.vainfo | default('NOT installed on this host — not a defect: Plex ships its own libva/media driver inside its container', true) }}
+
 --------------------------------------------------------------------------------
 PRESSURE STALL (/proc/pressure — % of wall time stalled)
 --------------------------------------------------------------------------------
@@ -52,7 +52,7 @@ All four steps are complete and each was verified, not assumed:
 | 1. Runner registered | ✅ `incluster-dind` online. Root cause of its absence was a stale `/data/.runner` identity; clearing the PVC forced re-registration |
 | 2. Gate observed green on a real PR | ✅ `gate / gate (pull_request)` ~1m35s |
 | 3. `enable_status_check: true` | ✅ context `gate / gate (pull_request)` — that one ONLY |
-| 4. Collaborator + approvals | ✅ `jakman85` is owner; `required_approvals: 1` |
+| 4. Collaborator + approvals | ✅ `jakman85` is owner. `required_approvals` was raised to 1 here and **lowered back to 0 on 2026-10-01** — see §6, it was a lock-out |

 Direct push to `main` was **provoked and rejected** (`pre-receive hook declined`) as
 `gitea-admin`, a site admin — so push protection is not merely declared, it holds for the
@@ -124,6 +124,66 @@ the stage goes **green on a real PR**, and only then set `REQUIRE_GITEA_SETTINGS
 **Changing the declaration is how you change policy.** If the live state is what you now want,
 edit this JSON in the same commit — deliberately, in review — rather than letting the two drift.

+### 6. `allow-admin-self-merge.sh` — undoing the lock-out step 4 created (2026-10-01)
+
+Step 4 above raised `required_approvals` to 1. That was a **lock-out, not a protection**, and the
+note at the bottom of `require-gate.sh` had already said why: *Gitea will not let an author approve
+their own PR.* With `jakman85` the only collaborator, there was no account in existence that could
+approve a PR `jakman85` opened.
+
+Measured 2026-10-01 — **PR #35**: `gate / gate (pull_request)` **success in 1m41s**, reviews `[]`,
+`required_approvals: 1`, `block_admin_merge_override: true`. Unmergeable, permanently, with no
+action available to its author. The same was true of any PR fixing it.
+
+It was also never the documented design. `ci/approval-policy.sh`'s header and `ci/gate.sh:194`
+both already describe the model as *"`required_approvals` drops to 0 and the GATE decides"*, and
+`decisions/2026-09-30-flux-deploy-notification-inhibited.md` flagged the 1-vs-0 disagreement as
+unresolved and not that session's to settle. So this is a correction toward the stated intent, not
+a weakening away from it.
+
+    ./infra/gitea/allow-admin-self-merge.sh      # required_approvals -> 0 (read from the declaration)
+
+**What stays switched on, and why that is the point.** The other way to unblock an owner is
+`block_admin_merge_override: false`, and it would have been the wrong one — that re-opens exactly
+what §5 above closed, letting the one account that merges bypass the gate itself. This script
+touches the approval **count** only and *verifies* afterwards that `block_admin_merge_override` and
+`enable_status_check` are both still `true`, failing loudly if the PATCH disturbed either:
+
+| | before | after |
+|---|---|---|
+| `required_approvals` | 1 | **0** |
+| `block_admin_merge_override` | true | true (asserted) |
+| `enable_status_check` | true | true (asserted) |
+
+So a **red gate is still unmergeable by everyone, administrators included**. What changed is that
+a *green* one no longer needs a signature that cannot exist.
+
+**Security-relevant changes are still reviewed** — by `ci/approval-policy.sh` against
+`approval-policy.json` `sensitive_paths`, not by this field. An administrator waives that review
+only by posting `/approve-override <reason>` on the PR: reason mandatory, permission re-checked
+live, invalidated by a new commit, and recorded by account in the gate summary.
+
+> **Post the override as a PR *review* (type "Comment"), not in the plain comment box** — until
+> the CI token is re-minted. `SETTINGS_API_TOKEN` is `[read:repository, read:admin]`, and
+> `/issues/{n}/comments` needs Gitea's *issue* scope, so Gitea answers with an error **object**
+> instead of an array and the override is invisible. Measured on PR #38 (2026-10-02): a correct,
+> fresh, well-formed override in the comment box produced `approval-policy: FAIL` reading as "no
+> approving review" — a permission failure disguised as a missing one. `/pulls/{n}/reviews` is
+> under `read:repository` and worked in the same run. The stage now reads **both** sources, names
+> which one it used, and says plainly when the comments API is unreadable. `set-ci-token.sh`
+> requests `read:issue` as of 2026-10-02, so **re-run it** and the comment box works too — token
+> scopes cannot be widened in place. Note that the
+**agent account `gitea-admin` also reports permission `owner`**, so it can clear its own PRs that
+way; that is an accepted consequence of authorising by permission alone, and the route to
+narrowing it (an `override_admins` allowlist, already honoured by the code) is written down in
+`approval-policy.json`.
+
+**Sequencing.** `ci/gitea-settings.sh` compares live against the declaration *from the branch under
+test*, and `REQUIRE_GITEA_SETTINGS=true`. So merge any other open PRs **before** running this, then
+run it, then re-run the gate on the branch carrying the new declaration. Any PR left open across
+the change declares 1 against a live 0 and goes **red until rebased** — the detector working, not
+a fault. The script's header spells the order out.
+
 ## `conformance-trigger-kubeconfig.sh` — the narrow L2/L3 trigger credential

 Implements `docs/plans/l2-l3-conformance-ci-wiring.md` §3 item 6 (= §5 step 7): mint a
@@ -204,6 +264,8 @@ it. The name matters — see the note under the behaviour table above.)
 2. **Watch the gate actually report on a real PR**, and note the context name it produced.
 3. **Only then set `enable_status_check: true`** with that exact context.
 4. Add the human reviewer as a repo collaborator, then raise `required_approvals` to 1.
+   — **Step 4 was REVERSED on 2026-10-01 (§6).** It assumed a reviewer distinct from the author
+   would exist. One never did, so this step produced a lock-out. Do not re-apply it.

 > **Do not do step 3 before step 1.** A required status check with no runner leaves every
 > PR permanently pending — including the PR that would fix it. That half is measured: PR #1
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# Let the sole administrator merge their own green PR: required_approvals 1 -> 0 on `main`.
+#
+# WHY. `required_approvals: 1` plus Gitea's refusal to let an author approve their own pull
+# request plus exactly ONE collaborator (`jakman85`, permission `owner`) is a lock-out, not a
+# policy. Every PR the operator opened was permanently unmergeable, including the PR that would
+# fix it. Measured 2026-10-01: PR #35 — `gate / gate (pull_request)` success in 1m41s, zero
+# reviews, `required_approvals: 1`, no account in existence able to satisfy it.
+#
+# The approvals field was also never what the design said it was. `ci/approval-policy.sh`'s
+# header and `ci/gate.sh:194` both already described the model as "required_approvals: 0 and the
+# GATE decides", and `require-gate.sh`'s own closing note named this exact failure and this exact
+# remedy. decisions/2026-09-30-flux-deploy-notification-inhibited.md flagged the 1-vs-0
+# disagreement as unresolved. This resolves it at 0.
+#
+# WHAT THIS DOES *NOT* DO, and the distinction is the whole point.
+#
+#   * It does NOT touch `block_admin_merge_override`, which stays TRUE. The required status check
+#     still binds ABSOLUTELY — a red gate is unmergeable by everyone, administrators included.
+#     Setting `block_admin_merge_override: false` would have been the other way to unblock an
+#     owner, and it would have re-opened precisely what block-admin-override.sh was written to
+#     close: PRs #5-#9 merged with zero approvals while the settings read as enforced.
+#   * It does NOT stop reviewing security-relevant changes. Those are gated by
+#     ci/approval-policy.sh against infra/gitea/approval-policy.json `sensitive_paths`, which an
+#     administrator can waive only with an explicit, attributed `/approve-override <reason>`
+#     comment on the PR.
+#
+# ORDER MATTERS — READ THIS BEFORE RUNNING IT.
+#
+# ci/gitea-settings.sh fails the merge gate on ANY divergence between the live rule and
+# infra/gitea/expected-branch-protection.json, and `gate.yml` sets REQUIRE_GITEA_SETTINGS=true.
+# The declaration in git now says 0. So between running this and merging that declaration, the
+# live value and the declared value must be compared from a branch that CARRIES the new
+# declaration — otherwise the stage goes red, correctly, for drift.
+#
+#   1. MERGE any other open PRs FIRST, while approvals still work for them (their branches
+#      declare 1, and a PR you did not author can still be approved normally). Skipping this
+#      only costs a rebase later; it does not break anything.
+#   2. Run this script.
+#   3. Re-run the gate on the branch that carries the new declaration. `gitea-settings` goes
+#      green because live 0 == declared 0.
+#   4. Merge that branch.
+#   5. Any PR still open from before step 2 now declares 1 against a live 0 and will be RED
+#      until it is rebased onto the new main. That is the detector doing its job, not a fault.
+#
+# Recoverable: `required_approvals` can be PATCHed back to 1 by an owner at any time. If you do,
+# change expected-branch-protection.json in the same breath or every PR goes red on drift.
+set -euo pipefail
+HOST=gitea.home.keithjackson.dev; OWNER=Home-Network; REPO=home-network-k3s-impl
+DECL="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/expected-branch-protection.json"
+
+command -v jq >/dev/null || { echo "jq is required" >&2; exit 1; }
+[ -f "$DECL" ] || { echo "declaration missing: ${DECL}" >&2; exit 1; }
+
+# Take the target from the DECLARATION, not from a literal here. If the two could disagree,
+# this script would be one more thing re-asserting a value nothing checks — which is the
+# failure mode the whole of infra/gitea/ exists to stop.
+WANT="$(jq -r '.protection.required_approvals' "$DECL")"
+[[ "$WANT" =~ ^[0-9]+$ ]] || { echo "declaration has no numeric required_approvals" >&2; exit 1; }
+
+creds="$(printf 'protocol=https\nhost=%s\n\n' "$HOST" | git credential fill)"
+U="$(sed -n 's/^username=//p' <<<"$creds")"; P="$(sed -n 's/^password=//p' <<<"$creds")"
+[ -n "$U" ] && [ -n "$P" ] || { echo "no stored credential for $HOST" >&2; exit 1; }
+echo "authenticating as '${U}' (secret not shown)"
+api() { curl -sS --config <(printf 'user = "%s:%s"\n' "$U" "$P") "$@"; }
+B="https://${HOST}/api/v1/repos/${OWNER}/${REPO}/branch_protections/main"
+
+# Read the RULE OBJECT, never /branches/main: the branch view does not expose
+# block_admin_merge_override and showed all-green for hours while the rule was overridable.
+echo "==> before"
+api "$B" | jq -r '{required_approvals, block_admin_merge_override, enable_status_check, status_check_contexts}'
+
+echo "==> setting required_approvals = ${WANT} (declared in expected-branch-protection.json)"
+api -X PATCH -H 'Content-Type: application/json' \
+  -d "{\"required_approvals\": ${WANT}}" "$B" >/dev/null
+
+echo "==> after"
+AFTER_JSON="$(api "$B")"
+jq -r '{required_approvals, block_admin_merge_override, enable_status_check, status_check_contexts}' <<<"$AFTER_JSON"
+
+GOT="$(jq -r '.required_approvals' <<<"$AFTER_JSON")"
+ADMIN_BLOCK="$(jq -r '.block_admin_merge_override' <<<"$AFTER_JSON")"
+STATUS="$(jq -r '.enable_status_check' <<<"$AFTER_JSON")"
+
+fail=0
+[ "$GOT" = "$WANT" ] || { echo "PATCH did not take: required_approvals is ${GOT}, wanted ${WANT}" >&2; fail=1; }
+# These two are what keeps this change narrow. If the PATCH disturbed either, the gate has
+# stopped binding and that must be loud, not discovered later.
+[ "$ADMIN_BLOCK" = true ] || { echo "block_admin_merge_override is ${ADMIN_BLOCK}, MUST be true — the gate is now bypassable" >&2; fail=1; }
+[ "$STATUS" = true ]      || { echo "enable_status_check is ${STATUS}, MUST be true — the gate is no longer required" >&2; fail=1; }
+[ "$fail" -eq 0 ] || { echo "do NOT record this as applied." >&2; exit 1; }
+
+cat <<'NOTE'
+
+NOW PROVE IT -- the setting is not the evidence. Two things to see, not assume:
+
+  1. An APPROVAL IS NO LONGER NEEDED. On your own PR with a green gate, the merge button is
+     live with zero reviews on it.
+  2. THE GATE STILL BINDS. On a PR whose gate is RED or still pending, Gitea must REFUSE the
+     merge -- for you, as owner. If it lets you through, block_admin_merge_override is not
+     doing its job and this change has traded a lock-out for a bypass.
+
+Then confirm the detector agrees: run ci/gate.sh from a branch carrying the new
+expected-branch-protection.json and check that `gitea-settings` is green rather than DRIFT.
+NOTE
@@ -1,6 +1,6 @@
 {
  "_intent": "The CI/CD pipeline is the approver. A human approval is required ONLY for changes that alter what the pipeline itself verifies, or the security posture it rests on -- because those are the changes CI cannot honestly gate: a PR that weakens the gate would be approved by the weakened gate.",
-  "_rule": "If a PR touches any sensitive_paths glob, it needs an APPROVED review from someone who is not its author. Everything else merges on a green gate alone.",
+  "_rule": "If a PR touches any sensitive_paths glob, it needs an APPROVED review from someone who is not its author, OR an admin override (see .override). Everything else merges on a green gate alone.",
  "_on_failure_to_evaluate": "BLOCK. Unlike gitea-settings, a skip here is a bypass, not a gap.",
  "sensitive_paths": [
    "ci/**",
@@ -24,5 +24,18 @@
    "platform/kyverno/**": "the admission policy engine",
    ".sops.yaml, **/*.sops.yaml, **/rbac.yaml": "encryption recipients, secrets, and cluster authority",
    "CLAUDE.md": "the delivery contract agents follow"
+  },
+  "override": {
+    "_why": "This repository has ONE human. Gitea refuses to let an author approve their own pull request, so 'a non-author APPROVED review' is a requirement the sole administrator can never satisfy -- every sensitive-path PR they open was permanently unmergeable, including the PR that would fix it. Measured 2026-10-01: PR #35 green gate, zero reviews, unmergeable. An escape hatch that the administrator can actually reach is therefore REQUIRED for the repository to function; the design question is only whether it leaves a record. This one does.",
+    "enabled": true,
+    "comment_directive": "/approve-override",
+    "requires_reason": true,
+    "authorized_permissions": ["admin", "owner"],
+    "invalidated_by_new_commits": true,
+    "_how": "Post the directive at the START of a line, followed by a reason: '/approve-override <why>'. ci/approval-policy.sh resolves the poster's LIVE repo permission via /collaborators/{login}/permission and accepts the override only if it is in authorized_permissions. The overrider, the source, the timestamp and the reason are printed in the gate log and the job summary.",
+    "_where_to_post_it": "EITHER a PR REVIEW body (type 'Comment') OR the ordinary PR comment box. Prefer the REVIEW: measured 2026-10-02 on PR #38, a well-formed override in the comment box was INVISIBLE to CI, because /issues/{n}/comments sits behind Gitea's *issue* token scope and SETTINGS_API_TOKEN is minted [read:repository, read:admin]. Gitea returned an error object rather than an array, and the stage reported 'no approving review' -- a permission problem wearing the costume of a missing override. /pulls/{n}/reviews is under read:repository and was readable in the same run, so a review body is the route that works on today's token. infra/gitea/set-ci-token.sh now also requests read:issue, which makes the comment box work -- but only after that script is RE-RUN, since an existing token's scopes cannot be widened in place. The stage accepts both sources and names which one it used, so neither route is load-bearing alone.",
+    "_self_override_is_allowed": "The author may override their own PR if they hold an authorized permission. That is the entire point -- with one human there is no one else -- but it means the 'a change to the gate cannot be gated by the gate' protection is, for an administrator, advisory rather than structural. It is traded for an auditable record: the waiver is an explicit act, attributed, with a stated reason, in the PR's comment history and the gate output.",
+    "_known_consequence_agent_accounts": "ACCEPTED, EYES OPEN (operator decision, 2026-10-01). Authorisation is by live permission ALONE, with no allowlist of human accounts. The agent account `gitea-admin` reports permission `owner` on this repo (measured 2026-10-01), so an AGENT can clear the sensitive-path review on its own PR by posting the directive. Agents author most PRs here, so this is the broad case, not the edge case. To narrow it later without changing any logic, add an `override_admins` allowlist here and require membership in addition to the permission check -- ci/approval-policy.sh reads both from this file.",
+    "_staleness": "An override is dropped unless its comment is NEWER than the PR head commit, mirroring Gitea's own dismiss_stale_approvals. An override earned at commit A therefore does not silently cover commit B; push again and you must override again."
  }
 }
@@ -1,6 +1,9 @@
 {
  "_comment": "DECLARED desired state of Gitea branch protection on main. ci/gitea-settings.sh fails the merge gate when the LIVE Gitea diverges from this. Gitea settings live in Gitea database - invisible to Flux, unreadable by every other check - so this file plus that check are the only drift detection that exists for them.",
  "_generated": "2026-09-29 from the live rule, after the session that found protection OFF and zero runners for a month",
+  "_amended": "2026-10-01: required_approvals 1 -> 0. See _why_required_approvals_is_zero.",
+  "_why_required_approvals_is_zero": "The GATE is the approver (operator intent, 2026-09-29). required_approvals: 1 was never consistent with that intent: ci/approval-policy.sh's header and ci/gate.sh:194 both already described the model as 'required_approvals: 0 and the GATE decides', and infra/gitea/require-gate.sh's own closing note named the consequence of 1 -- 'Gitea will not let an author approve their own PR'. With exactly ONE collaborator (jakman85, owner), that made every PR the operator opened permanently unmergeable. Measured 2026-10-01: PR #35, gate green in 1m41s, zero reviews, no path to a merge. decisions/2026-09-30-flux-deploy-notification-inhibited.md had already flagged the 1-vs-0 disagreement as unresolved; this settles it at 0.",
+  "_what_still_holds_at_zero": "This drops the APPROVAL COUNT only. block_admin_merge_override stays TRUE and enable_status_check stays TRUE, so the required gate context still binds absolutely -- nobody, administrator included, can merge a red gate. That distinction is the whole reason this is the chosen fix rather than block_admin_merge_override: false, which would have handed admins a bypass over the gate itself and re-opened what infra/gitea/block-admin-override.sh was written to close (PRs #5-#9, merged with zero approvals). Architecturally and security-relevant paths are still reviewed, by ci/approval-policy.sh rather than by this field.",
  "branch": "main",
  "runners_online_min": 1,
  "protection": {
@@ -11,7 +14,7 @@
    "status_check_contexts": [
      "gate / gate (pull_request)"
    ],
-    "required_approvals": 1,
+    "required_approvals": 0,
    "enable_approvals_whitelist": false,
    "block_admin_merge_override": true,
    "block_on_rejected_reviews": true,
@@ -34,7 +34,7 @@ jq -n '{
  enable_push: false,
  enable_status_check: true,
  status_check_contexts: ["gate / gate (pull_request)"],
-  required_approvals: 1,
+  required_approvals: 0,
  block_on_rejected_reviews: true,
  dismiss_stale_approvals: true,
  block_on_outdated_branch: false,
@@ -48,11 +48,21 @@ api "https://${HOST}/api/v1/repos/${OWNER}/${REPO}/branches/main" \

 cat <<'NOTE'

-NOTE ON required_approvals: 1 -- Gitea will not let an author approve their own PR.
-  * Agent-opened PRs (author gitea-admin) are approved by jakman85. Fine.
-  * PRs YOU open as jakman85 need gitea-admin to approve. If that is awkward, either open
-    agent work as gitea-admin and review as jakman85, or set required_approvals back to 0
-    and rely on the status check alone.
+NOTE ON required_approvals: 0 -- CHANGED FROM 1 ON 2026-10-01, and this script is the reason
+the change had to land here too: re-running it with 1 would silently re-impose the lock-out.
+
+The note that used to sit here called the problem and its own fix: "Gitea will not let an
+author approve their own PR ... if that is awkward ... set required_approvals back to 0 and
+rely on the status check alone." It was awkward. With jakman85 the ONLY collaborator, every PR
+the operator opened was unmergeable -- measured on PR #35, gate green in 1m41s, zero reviews,
+nothing that could ever satisfy the rule. So: 0, and the status check alone, as written.
+
+What did NOT change, deliberately:
+  * block_admin_merge_override stays TRUE. This drops the approval COUNT; it does not hand
+    anyone a bypass over the gate. A red gate is still unmergeable by everybody.
+  * Architecturally/security-relevant paths are still human-reviewed -- by ci/approval-policy.sh
+    (sensitive_paths in infra/gitea/approval-policy.json), which an administrator can waive
+    only by posting an explicit, attributed `/approve-override <reason>` comment.

 Then prove it, rather than trusting the API:
  git commit --allow-empty -m probe && git push origin main   # EXPECT: rejected
@@ -5,9 +5,23 @@
 # The token value is NEVER printed. It goes from the token API into the secrets API and
 # nowhere else -- not to stdout, not to a file, not into an agent transcript.
 #
-# Scope: `read:repository` + `read:admin` are what the check needs (the rule object, and the
-# runner list). Gitea token scopes are coarse; grant the least that works on your version and
-# confirm the check goes green rather than assuming the scope name.
+# Scope: `read:repository` + `read:admin` are what ci/gitea-settings.sh needs (the rule object,
+# and the runner list). `read:issue` is for ci/approval-policy.sh, and it was MISSING — see below.
+# Gitea token scopes are coarse; grant the least that works on your version and confirm the check
+# goes green rather than assuming the scope name.
+#
+# WHY read:issue IS HERE (added 2026-10-02, measured on PR #38). ci/approval-policy.sh reads the
+# PR's comments to find an admin `/approve-override`. /issues/{n}/comments sits behind the *issue*
+# scope, which this script did not request — so Gitea answered with an error OBJECT instead of an
+# array, and a perfectly well-formed override posted in the PR comment box was invisible to CI.
+# The failure did not look like a permission problem from the outside; it looked like the override
+# had not been posted. `/pulls/{n}/files` and `/pulls/{n}/reviews` are under read:repository, which
+# is why every other part of the same stage worked and this one did not.
+#
+# The stage now accepts an override from a REVIEW body as well, so it still functions on a token
+# WITHOUT read:issue. This scope is what makes the ordinary comment box work too — the thing
+# anybody would reach for first. **Re-run this script to pick it up; an existing token's scopes
+# cannot be widened in place.**
 set -euo pipefail
 HOST=gitea.home.keithjackson.dev; OWNER=Home-Network; REPO=home-network-k3s-impl
 # NOT GITEA_* — Gitea reserves that prefix for secret names and returns HTTP 400
@@ -21,7 +35,7 @@ basic() { curl -sS --config <(printf 'user = "%s:%s"\n' "$U" "$P") "$@"; }

 echo "==> minting token '${NAME}' as ${U}"
 TOKEN="$(basic -X POST -H 'Content-Type: application/json' \
-  -d "{\"name\":\"${NAME}\",\"scopes\":[\"read:repository\",\"read:admin\"]}" \
+  -d "{\"name\":\"${NAME}\",\"scopes\":[\"read:repository\",\"read:admin\",\"read:issue\"]}" \
  "https://${HOST}/api/v1/users/${U}/tokens" | jq -r '.sha1 // empty')"
 [ -n "$TOKEN" ] || { echo "token creation failed (check scopes for your Gitea version)" >&2; exit 1; }
 echo "==> got a ${#TOKEN}-character token (value not shown)"
@@ -106,14 +106,33 @@ spec:
    # CiliumL2AnnouncementPolicy (reserved 10.73.0.240-.250) authored by WS-NET in Phase 2.
    l2announcements:
      enabled: true
-    # --- cgroup: k3s ships its own cgroup mount; point Cilium at it, don't auto-mount ---
-    # [parity] MUST equal Ansible cgroup block. Not in the deliverable's headline list, but it IS
-    # part of the working CNI config on k3s: reconciling without it lets Cilium try to auto-mount
-    # cgroupv2 and break on adopt. Kept in parity for a safe Phase-2 adoption.
+    # --- cgroup: point Cilium at the HOST's cgroup2 root, which systemd mounts at /sys/fs/cgroup ---
+    # [parity] MUST equal Ansible cgroup block (roles/cilium/templates/cilium-values.yaml.j2).
+    #
+    # FIXED 2026-10-03 — hostRoot WAS /run/cilium/cgroupv2, AND THAT SILENTLY BROKE SOCKET-LB FOR
+    # EVERY HOST PROCESS. With autoMount disabled, the chart only bind-mounts whatever already exists
+    # at hostRoot ("it's expected that users have mounted cgroup2 filesystem at the specified
+    # cgroup.hostRoot" — cilium 1.20.1 values.yaml). Nothing ever mounted cgroup2 at
+    # /run/cilium/cgroupv2 on these nodes (it is an empty DirectoryOrCreate hostPath), so the agent
+    # mounted one ITSELF from inside its own private cgroup namespace. That mount's root is the
+    # agent container's OWN leaf cgroup, so cil_sock4_connect & co. were attached there and covered
+    # only the agent. Every other host-netns process (the k3s apiserver in
+    # /system.slice/k3s.service, curl over SSH) was untranslated: Service ClusterIPs timed out from
+    # every node, and the node holding the Gateway's L2 lease could not reach the Gateway VIP at all
+    # ("no route to host"). That is why k3s-prod-1's apiserver logged `oidc: authenticator not
+    # initialized` from 2026-09-12 on — kubectl SSO through the kube-vip VIP failed outright.
+    # Measured on k3s-prod-1 before this change: TCP connect to 192.168.2.20:443, 10.43.225.79:443
+    # and 10.43.59.141:8080 SUCCEEDED from inside the agent container (the covered cgroup) and
+    # FAILED from the host, same node, same netns. Full write-up:
+    # decisions/2026-10-03-cilium-cgroup-hostroot.md.
+    #
+    # /sys/fs/cgroup IS a cgroup2 mount in the host's root cgroup namespace on every node (verified
+    # with findmnt on all three), so bind-mounting it gives the agent the real host root. autoMount
+    # stays off because the mount it would create already exists.
    cgroup:
      autoMount:
        enabled: false
-      hostRoot: /run/cilium/cgroupv2
+      hostRoot: /sys/fs/cgroup
    # --- Agent rollout: ONE node at a time (blast-radius bound) — 2026-09-01 incident ---
    # [parity] MUST equal Ansible: updateStrategy.rollingUpdate.maxUnavailable: 1.
    # The chart's own default is 2 (cilium 1.20.1 values.yaml: updateStrategy.rollingUpdate.
@@ -41,11 +41,13 @@ hubble:
 # [parity] ingress VIP announcement on the flat LAN (docs/09, OQ-1)
 l2announcements:
  enabled: true
-# [parity] k3s provides the cgroupv2 mount; do not auto-mount
+# [parity] the host's cgroup2 root is /sys/fs/cgroup (systemd mounts it); do not auto-mount.
+# NOT /run/cilium/cgroupv2: nothing mounts cgroup2 there, so socket-LB attached to the agent's own
+# cgroup and host processes lost Service/LB translation (2026-10-03 — see helmrelease.yaml).
 cgroup:
  autoMount:
    enabled: false
-  hostRoot: /run/cilium/cgroupv2
+  hostRoot: /sys/fs/cgroup
 # [parity] roll the agent DaemonSet ONE node at a time. Overrides the cilium 1.20.1 chart
 # default of 2, which on 3 nodes lets a single bad render break 2/3 agents at once — as the
 # 2026-09-01 incident did (k8sServiceHost: "" → the `config` init container fell back to the
@@ -20,6 +20,30 @@ A wildcard `*.keithjackson.dev` matches only one label deep, so `<svc>.home.keit
 its own record. **`internal` services need NO public record** — PiHole resolves them to the internal
 Gateway VIP (split-horizon). ExternalDNS manages **`public`** services' records only.

+## Public target + in-cluster DDNS (2026-10-03)
+
+Gateway-published records are **CNAMEs to `wan.home.keithjackson.dev`**, not the Gateway VIP:
+
+```
+plex.home.keithjackson.dev  CNAME  wan.home.keithjackson.dev   <- external-dns (Gateway `target` annotation)
+wan.home.keithjackson.dev   A      <current WAN IP>            <- cloudflare-ddns (ddns.yaml), every 5m
+```
+
+- The `target` annotation is on the **shared Gateway** (`platform/gateway/gateway.yaml`), so it
+  applies to every route that publishes at all. `internal` routes (`controller: none`) still publish
+  nothing.
+- `ddns.yaml` runs `favonia/cloudflare-ddns` in this namespace with the same `cloudflare-api-token`.
+  It detects the WAN IP through Cloudflare's trace endpoint, so the pod **must not** carry
+  `egress-class: vpn`.
+- LAN and pods resolve `*.home.keithjackson.dev` to the Gateway VIP via PiHole and CoreDNS
+  (split-horizon) and never use the public path.
+- **Nodes resolve through public DNS** (`node_baseline_dns_servers`), so from a node, the apiserver's
+  OIDC issuer `keycloak.home.keithjackson.dev` resolves to the WAN IP. kubectl SSO therefore relies on
+  the router hairpinning (NAT reflection) WAN:443 back to the Gateway.
+
+Check: `dig +short @1.1.1.1 plex.home.keithjackson.dev` → `wan.home.keithjackson.dev.` then the WAN IP;
+`kubectl -n external-dns logs deploy/cloudflare-ddns`.
+
 ## Secrets

 Cloudflare API token referenced **by name** (`cloudflare-api-token`, shared with cert-manager),
@@ -0,0 +1,87 @@
+---
+# In-cluster dynamic DNS for the home WAN (2026-10-03).
+#
+# Keeps ONE record, `wan.${BASE_DOMAIN}` (A, IPv4 only, not proxied), pointed at the home WAN's
+# current public IP. Every public name external-dns publishes from the shared Gateway is a CNAME to
+# it (platform/gateway/gateway.yaml `external-dns.alpha.kubernetes.io/target`), so a WAN IP change
+# moves all of them through this one record, with no Git change and no external-dns involvement.
+#
+# OWNERSHIP SPLIT WITH EXTERNAL-DNS. This record carries no external-dns TXT registry entry, so
+# external-dns (policy: sync, owner `prod`) never treats it as its own and never deletes it.
+# Conversely this updater manages only the names in DOMAINS and never touches external-dns's records.
+#
+# IP DETECTION. Default `cloudflare.trace`: an HTTPS request to Cloudflare's trace endpoint, which
+# reports the address the request arrived from. Correct here only because this pod is NOT in the
+# `vpn` egress class — the CiliumEgressGatewayPolicy (platform/vpn-egress) selects pods by
+# `homelab.keithjackson.dev/egress-class: vpn` alone, cluster-wide. Labelling this pod `vpn` would
+# publish the VPN exit IP instead of the WAN IP.
+#
+# TOKEN. Reuses this namespace's `cloudflare-api-token` (same Zone:DNS:Edit token as external-dns
+# and cert-manager) — no new secret.
+#
+# TIER NOTE. This lives in the gating `platform` Flux Kustomization (wait: true) next to
+# external-dns. If it ever crashloops, `platform` goes not-Ready and apps stop reconciling (no
+# outage). Kept here rather than in its own non-gating Kustomization to avoid a clusters/** change.
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: cloudflare-ddns
+  namespace: external-dns
+  labels:
+    app.kubernetes.io/name: cloudflare-ddns
+spec:
+  replicas: 1                 # exactly one writer for the record
+  revisionHistoryLimit: 3
+  strategy:
+    type: Recreate            # never two updaters at once
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: cloudflare-ddns
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: cloudflare-ddns
+    spec:
+      automountServiceAccountToken: false
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        runAsGroup: 1000
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: cloudflare-ddns
+          image: favonia/cloudflare-ddns:1.17.1
+          imagePullPolicy: IfNotPresent
+          env:
+            - name: CLOUDFLARE_API_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: cloudflare-api-token
+                  key: api-token
+            - name: DOMAINS
+              value: "wan.${BASE_DOMAIN}"
+            - name: IP6_PROVIDER
+              value: "none"         # IPv4 WAN only; leaves any AAAA alone
+            - name: PROXIED
+              value: "false"        # DNS-only: Plex/Gitea/etc. terminate TLS on the Gateway
+            - name: UPDATE_CRON
+              value: "@every 5m"
+            - name: DELETE_ON_STOP
+              value: "false"        # a pod restart must not blank the record every CNAME depends on
+            - name: RECORD_COMMENT
+              value: "managed by k3s cloudflare-ddns (platform/externaldns/ddns.yaml)"
+            - name: EMOJI
+              value: "false"
+          resources:
+            requests:
+              cpu: 5m
+              memory: 16Mi
+            limits:
+              memory: 64Mi
+          securityContext:
+            allowPrivilegeEscalation: false
+            readOnlyRootFilesystem: true
+            capabilities:
+              drop:
+                - ALL
@@ -81,38 +81,29 @@ spec:
    # what this instance can see or own; per-resource exclusion only ever REMOVES an object from
    # this instance's view, so it cannot make prod capable of deleting a record it does not own.
    # ---------------------------------------------------------------------------------------------
-    # THE PUBLIC-TARGET GAP — STILL OPEN, AND NOT FIXABLE HERE ON v0.21.0 (operator decision)
+    # THE PUBLIC TARGET — DECIDED 2026-10-03 (operator): every Gateway-published name goes public
    #
-    # The five `exposure: public` services (plex, gitea, seerr, keycloak, nextcloud) publish records
-    # that point at 192.168.2.20 — the Gateway VIP, an unroutable RFC1918 address — so they are not
-    # actually reachable from the internet either. Their records are LEFT EXACTLY AS THEY ARE by
-    # this change, for two independent reasons:
+    # Until 2026-10-03 every gateway-sourced record answered with the Gateway VIP (192.168.2.20,
+    # RFC1918), so nothing published here was reachable from the internet. Two facts constrained the
+    # fix, and both still hold:
+    #   * `external-dns.alpha.kubernetes.io/target` is read in v0.21.0's gateway-httproute source ONLY
+    #     from the GATEWAY (source/gateway.go: `overrides: annotations.TargetsFromTargetAnnotation(
+    #     gw.Annotations)`) and from a ListenerSet listener — NEVER from an HTTPRoute. A target on the
+    #     shared Gateway therefore moves EVERY gateway-sourced record at once.
+    #   * The home WAN IP is residential and can change, so a literal IP in Git would rot.
    #
-    #   1. THE VALUE IS NOT ESTABLISHED. The one public IP this repo knows is 136.51.115.252
-    #      (clusters/prod/kustomization.yaml, where the zone's `*.keithjackson.dev` wildcard also
-    #      points). It belongs to the LEGACY non-k3s stack and currently serves a placeholder
-    #      certificate. Whether prod's public services belong behind it, behind a new forward, or
-    #      are not meant to be public at all is an operator decision. Breaking five services' DNS
-    #      on a guess is worse than leaving a known-wrong-but-stable state.
-    #   2. THE MECHANISM CANNOT BE SCOPED TO THEM. `external-dns.alpha.kubernetes.io/target` is a
-    #      real override, but in v0.21.0's gateway-httproute source it is read ONLY from the
-    #      GATEWAY (source/gateway.go: `overrides: annotations.TargetsFromTargetAnnotation(
-    #      gw.Annotations)`) and from a ListenerSet listener — NEVER from the HTTPRoute. So there is
-    #      no per-service target override while every public service shares the `shared` Gateway.
-    #      Annotating the Gateway would move EVERY gateway-sourced record, which after this change
-    #      still includes auth, crafty, grafana, headlamp and the apex — i.e. it would publish a
-    #      routable address for admin UIs. That is strictly worse than the status quo.
-    #
-    # SO THE DECISION THE OPERATOR OWNS IS TWO-PART, and no knob is added here until the first half
-    # is answered, because a knob that cannot take effect invites someone to set it and expect one:
-    #   (a) WHAT address should the five public names answer with — 136.51.115.252 as-is, a new
-    #       port-forward/WAN address, or nothing (drop `exposure: public` to `internal` instead)?
-    #   (b) HOW should they be separated from the non-public names that share the Gateway? Options,
-    #       roughly in ascending cost: a SECOND Gateway (or a ListenerSet) carrying the target
-    #       annotation, with only the public services' routes attached; or publishing the five names
-    #       from the `crd` source (DNSEndpoint objects the renderer emits for `public` services),
-    #       which restores per-service targeting without a second Gateway.
-    # Only (b) is a manifest change; (a) is a fact about the home WAN that has to be checked.
+    # The operator's answers: (a) yes, ALL of them — the five `exposure: public` apps AND the admin
+    # UIs (auth, crafty, grafana, headlamp, the apex), which sit behind Keycloak / oauth2-proxy;
+    # (b) the address is kept current IN-CLUSTER. Implementation:
+    #   * platform/gateway/gateway.yaml sets `target: wan.${BASE_DOMAIN}` on the shared Gateway, so
+    #     each published name becomes a CNAME to that one hostname;
+    #   * platform/externaldns/ddns.yaml (favonia/cloudflare-ddns) keeps `wan.${BASE_DOMAIN}` an A
+    #     record at the current WAN IP. That record has no external-dns TXT registry entry, so this
+    #     instance never owns or garbage-collects it.
+    # Internal services are unchanged: their routes carry `controller: none` (above), which is
+    # checked BEFORE the target is read, so they still publish nothing. The `service` source (the L4
+    # LoadBalancers crafty-mc, unifi) does not read the Gateway's annotation and still publishes its
+    # LB VIP. Whether those should go public is a separate, still-open question.
    # ---------------------------------------------------------------------------------------------
    # CROSS-CLUSTER DNS OWNERSHIP. Read before changing either knob below.
    #
@@ -8,3 +8,5 @@ resources:
  # SOPS-encrypted Cloudflare token (env CF_API_TOKEN) — SAME token as cert-manager, second copy in
  # the external-dns namespace. Flux decrypts at reconcile.
  - cloudflare-api-token.sops.yaml
+  # DDNS for `wan.${BASE_DOMAIN}` — the CNAME target every Gateway-published record points at.
+  - ddns.yaml
@@ -7,6 +7,23 @@ kind: Gateway
 metadata:
  name: shared
  namespace: gateway
+  annotations:
+    # PUBLIC DNS TARGET (2026-10-03, operator decision). Without this, external-dns publishes every
+    # route on this Gateway with the Gateway's OWN address (${GATEWAY_VIP}, RFC1918), which is
+    # unreachable from the internet. With it, every published `<svc>.${BASE_DOMAIN}` (and the apex)
+    # becomes a CNAME to `wan.${BASE_DOMAIN}`, an A record that platform/externaldns/ddns.yaml keeps
+    # pointed at the home WAN's current public IP. The WAN forwards 80/443 to ${GATEWAY_VIP}.
+    #
+    # Scope is EVERY route on this Gateway that publishes at all — the operator chose to make the
+    # admin UIs (auth, grafana, headlamp, crafty, apex) publicly resolvable alongside the five
+    # `exposure: public` apps. `internal` services stay unpublished via their
+    # `external-dns.alpha.kubernetes.io/controller: none` route annotation, which external-dns checks
+    # BEFORE it reads this target. external-dns v0.21.0 reads `target` from the Gateway only, never
+    # from an HTTPRoute — see platform/externaldns/helmrelease.yaml.
+    #
+    # LAN and in-cluster clients are unaffected: PiHole and CoreDNS answer ${BASE_DOMAIN} with
+    # ${GATEWAY_VIP} directly (split-horizon), so they never take this public path.
+    external-dns.alpha.kubernetes.io/target: "wan.${BASE_DOMAIN}"
 spec:
  gatewayClassName: cilium
  listeners:
@@ -0,0 +1,225 @@
+# platform/intel-gpu-plugin
+
+**Intel Device Plugin for Kubernetes (GPU)** — the standalone GPU plugin DaemonSet from
+[`intel/intel-device-plugins-for-kubernetes`](https://github.com/intel/intel-device-plugins-for-kubernetes),
+pinned at **v0.37.0**. It makes every node advertise **`gpu.intel.com/i915`** so a workload can
+*request* a GPU instead of hoping a `hostPath` is enough.
+
+- **Serves:** hardware transcoding for Plex (`hardware.gpu: intel-quicksync`), and any later
+  consumer of an Intel iGPU.
+- **Tier:** its **own** Flux Kustomization
+  ([`clusters/base/intel-gpu-plugin.yaml`](../../clusters/base/intel-gpu-plugin.yaml),
+  `dependsOn: cilium`, `wait: true`) — **not** a line in `platform/kustomization.yaml`. See
+  *Placement* below.
+- **Namespace:** `intel-gpu-plugin` — a PLATFORM namespace, labelled
+  `homelab.keithjackson.dev/app-namespace: "false"`, so the app-scoped Kyverno policies
+  (default-deny generation, auth-mode, egress-class, host-namespaces) do not apply.
+- **No RBAC, no ServiceAccount, no CRDs, no webhook, no API-server access.** A device plugin talks
+  to the *local kubelet* over a unix socket.
+
+## What this component owns
+
+| File | Purpose |
+|---|---|
+| `namespace.yaml` | `intel-gpu-plugin` ns, explicitly `app-namespace=false` |
+| `intel-gpu-plugin.yaml` | the DaemonSet, **vendored** (machine-generated) from upstream v0.37.0 |
+| `kustomization.yaml` | namespace scoping + the pinned image (`images:` transformer) |
+
+## The problem this fixes
+
+Plex was given a `/dev/dri` hostPath (merged on `main` @ `cdf3e14`). The device is **visible** in
+the container and **cannot be opened**. Measured in the running pod on `k3s-prod-2`:
+
+```
+uid 0,    O_RDWR -> EPERM "Operation not permitted"
+uid 1000, O_RDWR -> EPERM "Operation not permitted"
+```
+
+`EPERM` from the **device cgroup**, not `EACCES` from file permissions — proven three ways:
+
+1. root on the **host** opens the same device with `rc=0`;
+2. the node is unified `cgroup2fs`;
+3. the container's runc `config.json` reads
+   `linux.resources.devices: [{'allow': False, 'access': 'rwm'}]` and `linux.devices: None` —
+   one deny-all rule, no allow entry for major 226, and the runtime was never asked to grant a
+   device.
+
+A hostPath is a **filesystem** operation. It makes the device *node* appear and adds no cgroup
+allow rule, because Kubernetes cannot infer one from a volume.
+
+**A device plugin is the mechanism that closes exactly this gap.** Verified in the v0.37.0 source
+(`cmd/gpu_plugin/gpu_plugin.go`, `createDeviceSpecsFromDrmFiles`), the `Allocate` response carries
+per DRM node:
+
+```go
+specs = append(specs, pluginapi.DeviceSpec{
+    HostPath: devPath, ContainerPath: devPath, Permissions: "rw",
+})
+```
+
+and it is that `DeviceSpec` which makes kubelet populate `linux.devices` **and** the matching
+cgroup allow rule. It is the only **non-privileged** way to get there. `privileged: true` would
+also work and is not available: `platform/kyverno/policies/restrict-host-namespaces.yaml` blocks it
+under Enforce in app namespaces, and `plex` is labelled `app-namespace=true` (measured
+2026-10-01). Correctly so.
+
+## Placement — why its own Kustomization
+
+The cheap option was one line in the shared `platform` tier. Neither that file nor a new
+`platform/<name>/` directory is a sensitive path in `infra/gitea/approval-policy.json`, so it would
+have needed no human reviewer *on the policy's terms*. It is still the wrong place, and the repo
+already argued the case when Cilium was split out (read the comment in
+`platform/kustomization.yaml`):
+
+* `platform` is `wait: true` over eleven-plus components, so its Ready condition cannot distinguish
+  "the GPU plugin is wedged" from "longhorn is slow".
+* The only emergency valve would be `flux suspend kustomization platform` — freezing storage, TLS,
+  DNS, ingress, monitoring, Keycloak and KubeVirt in order to stop touching a device plugin.
+* The tier's 10m/5m interval/timeout are called out there as **wrong for a 3-node DaemonSet
+  rollout**. This is a 3-node DaemonSet rollout.
+
+The consequence is worse than average here because of what sits downstream: `platform` gates
+`platform-config`, which gates `apps-wave-0` (`wait: true`), which gates `apps-wave-1`. A plugin
+that never went healthy inside that tier would stall **every app** on "dependency not ready" — so a
+component added to fix a *silent no-op* (software transcoding) would have acquired the ability to
+cause a *real outage*.
+
+The usual objection — that `clusters/**` needs a human approval — costs nothing: Gitea enforces
+`required_approvals: 1` with `block_admin_merge_override: true` on **every** PR here regardless of
+path, and authors cannot self-approve. The approval was required either way.
+
+### `wait: true` is a health signal, not a gate
+
+This departs from the `headlamp`/`crafty` leaf template, deliberately. Those run `wait: false`
+because a leaf UI can be legitimately not-Ready for a long time during a cold start while it waits
+on Keycloak/DNS/TLS, making its health useless as a signal. This DaemonSet has no such dependency —
+it reads sysfs and opens a socket, so on healthy nodes it is Ready in seconds and **not-Ready
+genuinely means broken**. `wait: true` makes `flux get kustomization intel-gpu-plugin` a one-line
+answer to "did it come up on all three nodes", which is precisely the question that must be
+answered before the consumer PR may merge.
+
+Nothing in the graph `dependsOn` this, so `wait: true` gates nothing. **Think hard before adding
+`dependsOn: [intel-gpu-plugin]`** anywhere — for the app waves that would hand a device plugin the
+power to stall every app, which is the thing this placement exists to prevent.
+
+## Deliberate non-choices
+
+**Standalone DaemonSet, not the operator.** The operator adds a CRD (`GpuDevicePlugin`), a
+controller-manager Deployment, cert-manager-backed admission webhooks and a second namespace —
+three more things that can be unhealthy — to buy declarative management of a DaemonSet that is
+already declared in git. For three identical nodes and one consumer that is strictly more blast
+radius for no gain.
+
+**It lists devices; it never opens them** — worth being certain about, since *opening* is exactly
+what fails today. Verified in v0.37.0 source: the scan path uses only `os.ReadDir`, `os.ReadFile`,
+`os.Stat` and `os.Readlink` over `/sys/class/drm` and `/dev/dri`. There is no `os.Open`, no
+`OpenFile`, no `ioctl` and no raw syscall against a DRM node anywhere in
+`cmd/gpu_plugin/gpu_plugin.go`. Listing and `stat()` demonstrably work today — that is how the Plex
+probe found the device before failing to open it — so the plugin needs no cgroup grant of its own
+and no privilege. Accordingly `/dev/dri` and `/sys/class/drm` are mounted **readOnly**.
+
+* **`--health-management` stays OFF** (upstream default). It is the one flag that *would* need to
+  open a device: it starts a Level-Zero sidecar (`setupLevelZeroService` returns early unless the
+  flag is set). Enabling it would reintroduce the exact `EPERM` this component exists to fix.
+* `/dev/dri/by-path` being absent is non-fatal — `newDevicePlugin` warns and sets
+  `bypathFound = false`.
+
+**`--shared-dev-num` left at the upstream default of 1** (i.e. not set). The flag multiplies each
+card into N schedulable device IDs, letting N containers share one GPU. Not raised, because:
+
+* Plex is the only consumer today and wants one device;
+* the count is a **lie the scheduler believes** — there is no GPU-side partitioning or quota behind
+  it, so N containers get the same `renderD128` and contend for the same two media engines
+  (`vcs0`/`vecs0`). Raising it converts "Unschedulable", which is loud and obvious, into
+  "transcodes are mysteriously slow", which is neither;
+* it is a one-line, zero-migration change on the day a second consumer exists — and *then* the
+  right number will be known rather than guessed.
+
+**No NFD overlay.** All three nodes are identical (HP EliteDesk 800 G5 Mini, i7-9700T, UHD Graphics
+630, PCI `8086:3e98`, `i915` bound), so a nodeSelector on a GPU feature label would gate on
+something true everywhere. If a GPU-less node ever joins, add upstream's `nfd` overlay *then*
+rather than relying on the plugin's behaviour on such a node.
+
+**No `supplementalGroups`, anywhere.** The `pms-docker` image's own cont-init hook stats each
+`/dev/dri` char device and adds the `plex` user to its group — measured live as
+`Groups: 44 100 992 1000` on the running Plex process, where GID 992 is the host `render` group.
+That mechanism is image-side and works identically on a plugin-injected device.
+
+## Resource name
+
+`gpu.intel.com/i915`. Not a guess: `scan()` calls `devTree.AddDevice(devProps.driver(), ...)`, so
+the suffix is the **kernel driver name**. These nodes have `i915` bound (not `xe`), so the resource
+is `i915`. A future kernel moving CoffeeLake to `xe` would silently rename it to `gpu.intel.com/xe`
+and a consumer's request would stop matching — worth knowing before that kernel arrives.
+
+## Versioning
+
+Pinned to **v0.37.0** via the `images:` transformer in `kustomization.yaml`; the DaemonSet itself is
+vendored as `kubectl kustomize` output of the upstream base (see its header for the exact re-vendor
+command and both sha256s). It is *generated* rather than a byte copy because upstream mixes YAML
+sequence-indentation styles within the one document and `ci/yamllint.yaml` sets
+`indent-sequences: consistent` — a verbatim copy is a hard lint error that blocks the gate
+(measured: it did). `kubectl kustomize` normalises the formatting, and the result was proven to be
+a formatting-only change: `yq -P 'sort_keys(..)'` of the generated file and of the raw upstream file
+diff clean. This is also the re-vendor idiom `platform/external-snapshotter` documents.
+
+Vendored rather than referenced as a remote kustomize base for the same reason as
+`platform/external-snapshotter`: a remote base is `git fetch`ed on **every** reconcile, those
+fetches intermittently exceeded kustomize-controller's 27s timeout, and transient GitHub slowness
+then rendered as "kustomize build failed". A local file makes the build offline and deterministic —
+which also means the CI gate's `kustomize build` proves the manifests rather than proving GitHub
+was up.
+
+Renovate tracks it through the inline-marker idiom (`# renovate: datasource=github-releases
+depName=...` immediately above the version literal), because `renovate.json` sets
+`"docker": {"enabled": false}` and `"kubernetes": {"enabled": false}` — an image reference in a
+plain manifest is tracked by nothing. **Caveat:** upstream *release* tags carry a `v` (`v0.37.0`)
+while *image* tags do not (`0.37.0`). Whether Renovate normalises that for this dep is unverified;
+if it ever proposes `newTag: v0.38.0`, the fix is an `extractVersion` packageRule in
+`renovate.json`, not a hand-edit.
+
+## Ordering — do not reorder the stack
+
+The renderer change (`catalog/render.py` emitting `gpu.intel.com/i915: 1` for
+`hardware.gpu: intel-quicksync` instead of the hostPath) **must not land until this plugin is
+merged AND confirmed advertising the resource on all three nodes.** If the renderer requests a
+resource no node advertises, the Plex pod is **Unschedulable**; `apps-wave-0` is `wait: true` and
+`apps-wave-1` `dependsOn` it, so that stalls the whole app tier — turning today's silent no-op into
+a real outage.
+
+## Post-merge verification
+
+Nothing about this component's runtime behaviour can be verified from a PR: it does not exist until
+the operator merges and Flux reconciles. Afterwards:
+
+```bash
+D=./infra/devcontainer/dev.sh
+
+# 1. Flux applied it, and (wait:true) the DaemonSet went healthy.
+$D run bash -c 'flux get kustomization intel-gpu-plugin'
+$D run bash -c 'kubectl --context k3s-prod -n intel-gpu-plugin get ds intel-gpu-plugin'
+#    want: DESIRED=3 CURRENT=3 READY=3 AVAILABLE=3
+
+# 2. The resource is advertised on ALL THREE nodes. This is the gate on the renderer PR.
+$D run bash -c 'kubectl --context k3s-prod get nodes -o json \
+  | jq -r ".items[] | \"\(.metadata.name) i915=\(.status.allocatable[\"gpu.intel.com/i915\"] // \"ABSENT\")\""'
+#    want: three lines, each i915=1   (before this change: ABSENT on all three)
+
+# 3. If a node shows ABSENT, read that node's plugin logs.
+$D run bash -c 'kubectl --context k3s-prod -n intel-gpu-plugin logs -l app=intel-gpu-plugin --tail=50 --prefix'
+```
+
+### The decisive test
+
+Resource-advertised is necessary but **not** sufficient — it proves the plugin registered, not that
+the cgroup allow rule reaches Plex. The decisive test is re-running the `open()` probe **as uid 1000
+through `s6-setuidgid plex`** inside the Plex pod, *after* the renderer PR has switched Plex to
+request the resource:
+
+```bash
+$D run bash -c 'kubectl --context k3s-prod -n plex exec deploy/plex -- \
+  s6-setuidgid plex python3 -c "import os; os.close(os.open(\"/dev/dri/renderD128\", os.O_RDWR)); print(\"rc=0 OK\")"'
+```
+
+**Passing as uid 0 is weaker evidence and must not be accepted as proof** — Plex transcodes as uid
+1000, and the two differ in group membership, which is the half the image's cont-init hook supplies.
@@ -0,0 +1,117 @@
+---
+# VENDORED, MACHINE-GENERATED, from intel/intel-device-plugins-for-kubernetes @ v0.37.0:
+#   deployments/gpu_plugin/base
+#
+# RE-VENDOR on a version bump (and bump `newTag` in kustomization.yaml to match):
+#   kubectl kustomize \
+#     "https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/gpu_plugin/base?ref=vX.Y.Z" \
+#     > platform/intel-gpu-plugin/intel-gpu-plugin.yaml
+#   # then re-add this header, and READ THE DIFF — upstream has changed the securityContext
+#   # and the volume set between minor releases before.
+#
+# Generated 2026-10-01 at v0.37.0; sha256 of the body below (header excluded):
+#   80ebd070f983f3116f4004f719d9c872220997ec766b4ab7e7290585ac8ef842
+# The raw upstream source file reviewed to produce it
+# (deployments/gpu_plugin/base/intel-gpu-plugin.yaml) was sha256
+#   c8773918c2370e711d4ba84120a9971a290bf75ecf81dc520782bcccd7ef24b6
+# and the two were proven EQUIVALENT, not merely similar: `yq -P 'sort_keys(..)'` of each diffs
+# clean, so passing it through kustomize changed formatting only and no content.
+#
+# WHY GENERATED RATHER THAN A BYTE COPY of the upstream file: upstream mixes YAML sequence
+# indentation styles within the one document (`containers:` non-indented, `env:` indented), and
+# ci/yamllint.yaml sets `indentation: indent-sequences: consistent`, so a verbatim copy is a hard
+# lint ERROR and blocks the merge gate — measured, this file failed exactly that way first time
+# round. `kubectl kustomize` normalises the output, which is also precisely the re-vendor idiom
+# platform/external-snapshotter documents for the same reason.
+#
+# WHY VENDORED AT ALL rather than a remote kustomize base (`...gpu_plugin/base?ref=v0.37.0` left in
+# `resources:`): the same reason platform/external-snapshotter is vendored. A remote base is
+# `git fetch`ed by Flux's kustomize-controller on EVERY reconcile, those fetches intermittently
+# exceeded the controller's 27s timeout, and a transient GitHub slowness then rendered as
+# "kustomize build failed" on the whole tier. A local file makes the build OFFLINE and
+# deterministic — which also means `kustomize build` in the CI gate proves the real thing instead
+# of proving GitHub was up. (platform/kubevirt still carries the remote-base pattern and its latent
+# fragility; not this component's problem.)
+#
+# DO NOT HAND-EDIT THE BODY. Every local change belongs in kustomization.yaml as a transformer or
+# patch, so the next re-vendor stays a mechanical regenerate-and-diff instead of a three-way merge.
+# The image tag below is re-pinned (and fully qualified) by the `images:` transformer there. The
+# rationale for each thing deliberately NOT changed — no privileged, no supplementalGroups, no
+# -shared-dev-num, no NFD overlay, no health-management — is in README.md.
+apiVersion: apps/v1
+kind: DaemonSet
+metadata:
+  labels:
+    app: intel-gpu-plugin
+  name: intel-gpu-plugin
+spec:
+  selector:
+    matchLabels:
+      app: intel-gpu-plugin
+  template:
+    metadata:
+      labels:
+        app: intel-gpu-plugin
+    spec:
+      containers:
+      - env:
+        - name: NODE_NAME
+          valueFrom:
+            fieldRef:
+              fieldPath: spec.nodeName
+        - name: HOST_IP
+          valueFrom:
+            fieldRef:
+              fieldPath: status.hostIP
+        image: intel/intel-gpu-plugin:0.37.0
+        imagePullPolicy: IfNotPresent
+        name: intel-gpu-plugin
+        resources:
+          limits:
+            cpu: 100m
+            memory: 115Mi
+          requests:
+            cpu: 40m
+            memory: 45Mi
+        securityContext:
+          allowPrivilegeEscalation: false
+          capabilities:
+            drop:
+            - ALL
+          readOnlyRootFilesystem: true
+          seLinuxOptions:
+            type: container_device_plugin_t
+          seccompProfile:
+            type: RuntimeDefault
+        volumeMounts:
+        - mountPath: /dev/dri
+          name: devfs
+          readOnly: true
+        - mountPath: /sys/class/drm
+          name: sysfsdrm
+          readOnly: true
+        - mountPath: /var/lib/kubelet/device-plugins
+          name: kubeletsockets
+        - mountPath: /var/run/cdi
+          name: cdipath
+      nodeSelector:
+        kubernetes.io/arch: amd64
+      volumes:
+      - hostPath:
+          path: /dev/dri
+        name: devfs
+      - hostPath:
+          path: /sys/class/drm
+        name: sysfsdrm
+      - hostPath:
+          path: /var/lib/kubelet/device-plugins
+        name: kubeletsockets
+      - hostPath:
+          path: /var/run/cdi
+          type: DirectoryOrCreate
+        name: cdipath
+  updateStrategy:
+    rollingUpdate:
+      maxSurge: 0
+      maxUnavailable: 1
+    type: RollingUpdate
@@ -0,0 +1,99 @@
+---
+# Intel GPU device plugin for Kubernetes — makes the nodes advertise `gpu.intel.com/i915`.
+#
+# WHAT THIS FIXES, and why nothing cheaper works. Plex already has a `/dev/dri` hostPath mount
+# (merged on main @ cdf3e14). The device node is VISIBLE in the container and cannot be OPENED.
+# Measured in the running pod on k3s-prod-2:
+#     uid 0,    O_RDWR -> EPERM "Operation not permitted"
+#     uid 1000, O_RDWR -> EPERM "Operation not permitted"
+# EPERM from the DEVICE CGROUP, not EACCES from file permissions — proven three ways: root on the
+# HOST opens the same device rc=0; the node is unified cgroup2fs; and the container's runc
+# config.json carries one deny-all rule and nothing else:
+#     linux.resources.devices: [{'allow': False, 'access': 'rwm'}]
+#     linux.devices: None
+# No allow entry for major 226, and the runtime was never asked to grant a device. A hostPath is a
+# FILESYSTEM operation: it makes the device node appear, and adds no cgroup allow rule, because
+# Kubernetes cannot infer one from a volume.
+#
+# A device plugin is the mechanism that closes exactly that gap. Verified by reading the v0.37.0
+# source (cmd/gpu_plugin/gpu_plugin.go, createDeviceSpecsFromDrmFiles): the Allocate response
+# carries `pluginapi.DeviceSpec{HostPath, ContainerPath, Permissions: "rw"}` per DRM node, and it
+# is that DeviceSpec which makes kubelet populate `linux.devices` AND the matching cgroup allow
+# rule. It is the only NON-PRIVILEGED way to get there. `privileged: true` would also work and is
+# not available: platform/kyverno/policies/restrict-host-namespaces.yaml blocks it under Enforce in
+# app namespaces, and `plex` is labelled app-namespace=true (measured 2026-10-01). Correctly so.
+#
+# STANDALONE DAEMONSET, NOT THE OPERATOR. Upstream ships both. The operator adds a CRD
+# (GpuDevicePlugin), a controller-manager Deployment, cert-manager-backed admission webhooks and a
+# second namespace — i.e. three more things that can be unhealthy — to buy declarative management
+# of a DaemonSet we are declaring in git anyway. For three identical nodes and one consumer that is
+# strictly more blast radius for no gain. The DaemonSet has no CRDs, no controller, no webhook, and
+# no API-server access at all: a device plugin talks to the LOCAL kubelet over a unix socket, which
+# is why there is no ServiceAccount or RBAC here (and therefore no `**/rbac.yaml` sensitive path).
+#
+# LISTING, NOT OPENING — the thing worth being sure about, since opening is precisely what fails
+# today. Verified in v0.37.0 source: the scan path uses only os.ReadDir / os.ReadFile / os.Stat /
+# os.Readlink over /sys/class/drm and /dev/dri. There is no os.Open, no OpenFile, no ioctl and no
+# syscall of any kind against a DRM node anywhere in cmd/gpu_plugin/gpu_plugin.go. Listing and
+# stat() demonstrably work today (that is how the Plex probe found the device before failing to
+# open it), so the plugin needs no cgroup grant of its own and no privilege.
+#   * `--health-management` is left OFF (upstream default). It is the one flag that WOULD need to
+#     open a device: it starts a Level-Zero sidecar (setupLevelZeroService returns early unless the
+#     flag is set). Turning it on would reintroduce the exact EPERM this component exists to fix.
+#   * /dev/dri and /sys/class/drm are mounted readOnly, consistent with list-and-stat only.
+#   * /dev/dri/by-path absence is non-fatal (newDevicePlugin warns and sets bypathFound=false).
+#
+# RESOURCE NAME: `gpu.intel.com/i915`. Not a guess — `scan()` does
+# `devTree.AddDevice(devProps.driver(), ...)`, so the suffix is the KERNEL DRIVER name. These nodes
+# have i915 bound (not xe) on UHD Graphics 630 / PCI 8086:3e98, so the resource is i915. A future
+# kernel moving CoffeeLake to xe would silently rename it to `gpu.intel.com/xe` and the consumer's
+# request would stop matching — worth knowing before that kernel arrives.
+#
+# `--shared-dev-num` LEFT AT THE UPSTREAM DEFAULT OF 1 (so: not set). The flag multiplies each
+# card into N schedulable device IDs (`for i := 0; i < sharedDevNum; i++`), letting N containers
+# share one GPU. Deliberately not raised:
+#   * Plex is the only consumer and wants one device.
+#   * The count is a LIE the scheduler believes. There is no GPU-side partitioning or quota behind
+#     it — N containers get the same renderD128 and contend for the same two media engines
+#     (vcs0/vecs0). Raising it converts "Unschedulable", which is loud and obvious, into
+#     "transcodes are mysteriously slow", which is neither.
+#   * It is a one-line, zero-migration change on the day a second consumer actually exists, and at
+#     that point the right number will be known rather than guessed.
+# No NFD overlay either: all three nodes are identical (established cluster fact), so a nodeSelector
+# on a GPU feature label would gate on something that is true everywhere. If a GPU-less node ever
+# joins, add upstream's nfd overlay THEN rather than relying on the plugin's behaviour there.
+#
+# NO supplementalGroups, anywhere, on purpose. The pms-docker image's own cont-init hook stats each
+# /dev/dri char device and adds the `plex` user to its group — measured live as `Groups: 44 100 992
+# 1000` on the running Plex process, where GID 992 is the host `render` group. That mechanism is
+# image-side and works identically on a plugin-injected device, so adding a supplementalGroup here
+# or on the consumer would duplicate a thing that already works.
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+namespace: intel-gpu-plugin
+resources:
+  - namespace.yaml
+  - intel-gpu-plugin.yaml
+images:
+  # PINNED, never `latest`. Upstream release v0.37.0 (2026-09-16); verified present on Docker Hub
+  # 2026-10-01, manifest digest sha256:7001200071db10587fc31184eedda4f5cf291efba737adf85a8c80539f7fa5e1
+  # (single amd64 manifest, not an index). Recorded here rather than pinned by digest because this
+  # repo pins upstream platform components by VERSION TAG (platform/kubevirt, external-snapshotter);
+  # digest pinning is reserved for the conformance runner, where it is a security boundary.
+  # `newName` fully qualifies the registry: upstream writes the bare `intel/intel-gpu-plugin`, which
+  # relies on containerd's docker.io default rather than saying so.
+  #
+  # RENOVATE: the inline-marker idiom (customManagers in renovate.json), not the docker datasource —
+  # renovate.json sets `"docker": {"enabled": false}` and `"kubernetes": {"enabled": false}`, so an
+  # image reference in a plain manifest is tracked by NOTHING. The marker must sit on the line
+  # IMMEDIATELY above the version literal for the regex to match, which is why it is down here on
+  # `newTag` and not up in this header.
+  # CAVEAT, stated because it will bite someone: the GitHub release tags carry a `v`
+  # (`v0.37.0`) and the IMAGE tags do not (`0.37.0` — confirmed against the Docker Hub tag list).
+  # Whether Renovate normalises that for this dep is UNVERIFIED here (no Renovate run available).
+  # If it ever proposes `newTag: v0.38.0`, the fix is an `extractVersion` packageRule in
+  # renovate.json, not a hand-edit of this line.
+  - name: intel/intel-gpu-plugin
+    newName: docker.io/intel/intel-gpu-plugin
+    # renovate: datasource=github-releases depName=intel/intel-device-plugins-for-kubernetes
+    newTag: "0.37.0"
@@ -0,0 +1,34 @@
+---
+# intel-gpu-plugin — namespace for the Intel GPU device plugin DaemonSet.
+#
+# PLATFORM INFRA, NOT AN APPLICATION NAMESPACE, and that distinction is load-bearing here rather
+# than cosmetic. Every Kyverno policy that could object to this DaemonSet is scoped by a
+# `namespaceSelector: matchLabels: homelab.keithjackson.dev/app-namespace: "true"` — verified by
+# reading the policies, not assumed:
+#   * platform/kyverno/policies/restrict-host-namespaces.yaml  (hostNetwork/PID/IPC + privileged)
+#   * platform/kyverno/policies/require-egress-class-label.yaml
+#   * platform/kyverno/policies/require-auth-mode-label.yaml
+#   * platform/kyverno/policies/require-default-deny-baseline.yaml (a *generate* rule, so an
+#     app-namespace also gets a default-deny NetworkPolicy minted into it)
+# Labelling this namespace "false" therefore puts it out of scope for all four. That is the same
+# mechanism restrict-host-namespaces' own header describes: "platform components that do [need host
+# access] live in platform namespaces that do NOT carry the app-namespace label".
+#
+# Measured on prod 2026-10-01: `kyverno` and `vpn-egress` are the two namespaces that carry the
+# label explicitly as "false"; the rest of the platform namespaces leave it unset. Both are out of
+# scope identically — an unset label does not match `"true"`. We set it EXPLICITLY, following
+# vpn-egress, because "someone chose this" and "nobody thought about it" should not look the same
+# in git, and because this component's correctness depends on the answer.
+#
+# NOTE this plugin does NOT actually need any of the things those policies forbid: it sets no
+# hostNetwork/hostPID/hostIPC and is not privileged (see intel-gpu-plugin.yaml). The namespace
+# placement is not a workaround for a policy — there is nothing to work around. It is here because
+# a node-level platform DaemonSet is not an application, and because the app-namespace fabric
+# (default-deny NetworkPolicy, egress-class, auth-mode) describes obligations that are meaningless
+# for a process that talks only to the local kubelet over a unix socket and never uses the network.
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: intel-gpu-plugin
+  labels:
+    homelab.keithjackson.dev/app-namespace: "false"
@@ -88,3 +88,72 @@ This proves the R-IDP-02 property (unauthenticated cannot reach the upstream) on
 3. Log in as a **non-platform-admins** user -> expect **403** (authenticated but not authorized).
 4. Confirm the session cookie is `.home.keithjackson.dev`-scoped (single sign-on across the other
   protected UIs).
+
+## Observability: what this component can and cannot see (measured 2026-09-30)
+
+`servicemonitor.yaml` scrapes **this namespace's** deployment and nothing else. On prod that is
+two targets, `job="oauth2-proxy"`, both `up=1`. The complete set of oauth2-proxy-specific series
+in the TSDB is:
+
+    count by (__name__) ({__name__=~"oauth2_proxy.*"})
+      oauth2_proxy_requests_in_flight => 2
+
+**The per-app gate sidecars are not scraped, and there is nothing to scrape.** Six apps run an
+`oauth2-proxy` container of their own — crafty, jackett, pihole, qbittorrent, radarr, sonarr —
+each with `--http-address=0.0.0.0:4180` and **no `--metrics-address`**, so oauth2-proxy's metrics
+listener never starts. They declare only a `http:4180` container port, no `metrics` port exists on
+their Services, and `kubectl get servicemonitors,podmonitors -A` selects none of them. Any alert
+written against `oauth2_proxy_requests_total{namespace="jackett"}` would be structurally inert.
+
+This matters because the gate's characteristic failure is invisible by construction: an
+unauthenticated **XHR** gets a `302` to Keycloak rather than a `401`, the browser follows it
+against a live SSO session with no prompt, and the call resolves **200 carrying an HTML page**
+where the app expected JSON. Nothing in the request path reports an error. On 2026-09-30 jackett
+spent a burst of requests in exactly that state, re-established a session, and lost it 26 seconds
+into a 5-minute session, while `jackett-down` sat at 1 replica available the whole time.
+
+`auth-visibility-alerts.yaml` therefore alerts on **the gap**, not on the incident:
+`OAuth2ProxyGateUnobserved` compares `kube_pod_container_info{container="oauth2-proxy"}` against
+`up{container="oauth2-proxy"}` — two already-scraped families — and names every namespace running
+a gate that nothing observes. It fires for six namespaces today and clears per-namespace as each
+one gains a scrape. `runbooks/oauth2-proxy-auth-failures.md` is the by-hand substitute meanwhile,
+and it reads the sidecar logs out of Loki (Alloy collects them; 72h retention).
+
+**To close the gap for an app**, three things are needed together, on that app's own manifest:
+`--metrics-address=0.0.0.0:44180` on the sidecar, a `metrics` containerPort plus the matching
+Service port, and a ServiceMonitor selecting it. Only then does the auth-churn rule — sustained
+`code="302"` as a share of `oauth2_proxy_requests_total`, held across a window longer than the
+session lifetime — become writable rather than decorative.
+
+### Update 2026-10-01: the listener exists; the scrape path does not
+
+PR #28 added `--metrics-address=0.0.0.0:44180` and a named `metrics` container port to the five
+fleet sidecars (jackett, radarr, sonarr, qbittorrent, pihole). `podmonitor.yaml` in this directory
+discovers them by the pod label `homelab.keithjackson.dev/auth-mode: forward-auth` — a **PodMonitor**
+rather than a ServiceMonitor because no Service exposes 44180 and adding one would mean changing the
+renderer.
+
+The endpoint is real: port-forwarded straight to the jackett pod, `/metrics` returns **http=200,
+10208 bytes**, exporting `oauth2_proxy_requests_in_flight`, `oauth2_proxy_requests_total` (with
+`code` labels) and `oauth2_proxy_response_duration_seconds`. That is verified by payload, not by the
+flag appearing in `args` — a distinction worth keeping, because the two claims came apart here.
+
+**The scrape is nevertheless dropped.** From a pod in ns `monitoring`, both `:44180` and `:4180` on
+the jackett pod time out; the shared proxy's `:44180` answers. The only difference is that ns
+`oauth2-proxy` has **no NetworkPolicy at all**, while each app namespace has a `default-deny` whose
+`policyTypes` include `Ingress` with no `ingress` rules (deny-all) plus a single CNP admitting
+`fromEntities: [ingress]` to port 4180 — the Gateway's Envoy and nothing else. Prometheus matches no
+clause.
+
+So the targets read `up == 0` and the TSDB is still empty for those namespaces. This is the same bug
+and the same fingerprint as `platform/longhorn/networkpolicy-metrics.yaml` documents for Longhorn,
+and the same narrow fix applies: ns `monitoring` → the gate pod label → TCP 44180 only. It cannot
+ship from this directory, because `oauth2-proxy` and `apps-wave-0` are **siblings** in the Flux graph
+(both `dependsOn: [platform-config]`), so a NetworkPolicy naming ns `jackett` would apply on today's
+cluster and fail on a cold rebuild. It belongs with each app's own rendered policy.
+
+Consequence for `auth-visibility-alerts.yaml`: its `unless` clause requires `up == 1`, not merely a
+present `up` series. A discovered-but-failing target publishes `up` with value 0, so the bare form
+would have cleared `OAuth2ProxyGateUnobserved` the moment the PodMonitor landed — reporting the gates
+as observed while nothing was being scraped. The unit test pins that scenario; deleting the `== 1`
+fails it and nothing else.
@@ -0,0 +1,177 @@
+---
+# promtool unit tests for OAuth2ProxyGateUnobserved (platform/oauth2-proxy/auth-visibility-alerts.yaml).
+#
+# WHY. The rule is an `unless` between two aggregations over DIFFERENT metric families, and the
+# failure modes of that shape are both silent: a label that does not survive `count by` makes the
+# right-hand side match nothing (the alert fires for the scraped namespace too, i.e. permanently),
+# and a right-hand side that over-matches makes it fire for nobody (i.e. inert). Neither is
+# visible by reading the expression, and neither is visible on a live cluster where the answer
+# happens to look plausible. So all four states are pinned: fires for an unscraped namespace,
+# stays quiet for a scraped one, honours `for: 1h`, and STOPS once a scrape appears.
+#
+# HOW TO RUN. The one repo-wide harness, by the companion convention:
+#   ./infra/devcontainer/dev.sh run bash -c 'platform/monitoring/tests/run.sh platform/oauth2-proxy/auth-visibility-alerts.promtool-test.yaml'
+# promtool is pinned in infra/devcontainer/versions.env; never run it off the host PATH.
+#
+# THE SERIES BELOW ARE REAL SHAPES read off live prod Prometheus on 2026-09-30 before being
+# copied here — kube_pod_container_info as kube-state-metrics emits it (job="kube-state-metrics",
+# its own instance/pod labels, image/image_id/container_id) and `up` as the oauth2-proxy
+# ServiceMonitor emits it (endpoint="metrics", port 44180). The point of using the real shapes is
+# that `count by (namespace)` has to discard all of them; a test with tidy two-label series would
+# pass while the real expression collapsed differently.
+#
+# ASSERTIONS USE `ALERTS` RATHER THAN exp_alerts: the description annotation is long prose and
+# comparing it verbatim would couple this file to every wording change while proving nothing the
+# ALERTS series does not already prove.
+rule_files:
+  - auth-visibility-alerts.rules.yaml
+evaluation_interval: 1m
+
+tests:
+  # ---------------------------------------------------------------------------------------------
+  # 1. THE STATE PROD IS ACTUALLY IN (2026-09-30): jackett runs a gate sidecar with no scrape;
+  #    the shared oauth2-proxy namespace runs two that ARE scraped. Asserts the alert fires for
+  #    jackett and NOT for oauth2-proxy, and that `for: 1h` is honoured (pending at 59m).
+  - interval: 1m
+    name: an unscraped gate namespace fires after 1h; the scraped one never does
+    input_series:
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://937ae6fc",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="jackett",pod="jackett-c5bfcbfb8-c75f9",service="kube-prometheus-stack-kube-state-metrics",uid="5b57d48f-e7d3-4050-9cd1-16aaaa052996"}'
+        values: '1x130'
+      # The shared instance: kube-state-metrics sees both replicas...
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://aa11bb22",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="oauth2-proxy",pod="oauth2-proxy-7568c78795-cqz27",service="kube-prometheus-stack-kube-state-metrics",uid="1111aaaa-0000-0000-0000-000000000001"}'
+        values: '1x130'
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://cc33dd44",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="oauth2-proxy",pod="oauth2-proxy-7568c78795-bscxg",service="kube-prometheus-stack-kube-state-metrics",uid="1111aaaa-0000-0000-0000-000000000002"}'
+        values: '1x130'
+      # ...and both replicas are genuinely scraped, so `up` carries container="oauth2-proxy" there.
+      - series: 'up{container="oauth2-proxy",endpoint="metrics",instance="10.0.1.189:44180",job="oauth2-proxy",namespace="oauth2-proxy",pod="oauth2-proxy-7568c78795-cqz27",service="oauth2-proxy"}'
+        values: '1x130'
+      - series: 'up{container="oauth2-proxy",endpoint="metrics",instance="10.0.2.47:44180",job="oauth2-proxy",namespace="oauth2-proxy",pod="oauth2-proxy-7568c78795-bscxg",service="oauth2-proxy"}'
+        values: '1x130'
+    promql_expr_test:
+      # Not yet firing at 59m — `for: 1h` has not elapsed. Guards against someone "simplifying"
+      # the hold away and turning every rollout into a page.
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing"}'
+        eval_time: 59m
+        exp_samples: []
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="pending",namespace="jackett"}'
+        eval_time: 59m
+        exp_samples:
+          - labels: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="pending",namespace="jackett",severity="warning"}'
+            value: 1
+      # Firing for jackett past the hold...
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing"}'
+        eval_time: 65m
+        exp_samples:
+          - labels: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="jackett",severity="warning"}'
+            value: 1
+      # ...and never for the namespace that IS scraped. This is the assertion that catches an
+      # `unless` whose right-hand side stopped matching: without it, a broken rule that fires for
+      # every gate namespace would pass the test above.
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",namespace="oauth2-proxy"}'
+        eval_time: 65m
+        exp_samples: []
+      # The count must be 1 per namespace-with-a-gate, not 1 overall: a `count by (namespace)`
+      # that lost its `by` would collapse to a single unlabelled series.
+      - expr: 'count by (namespace) (kube_pod_container_info{container="oauth2-proxy"}) unless count by (namespace) (up{container="oauth2-proxy"})'
+        eval_time: 65m
+        exp_samples:
+          - labels: '{namespace="jackett"}'
+            value: 1
+
+  # ---------------------------------------------------------------------------------------------
+  # 2. THE FIX LANDING. jackett's sidecar gains --metrics-address and a ServiceMonitor 70m in, so
+  #    `up` starts carrying container="oauth2-proxy" for namespace="jackett" under a DIFFERENT job
+  #    name than the shared proxy's. The alert must stop. This is what makes the rule a gap
+  #    tracker rather than a permanent nag, and it pins that the match is on `container`, not on
+  #    a hardcoded job="oauth2-proxy" (which would keep firing forever after a correct fix).
+  - interval: 1m
+    name: the alert clears once the sidecar is actually scraped
+    input_series:
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://937ae6fc",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="jackett",pod="jackett-c5bfcbfb8-c75f9",service="kube-prometheus-stack-kube-state-metrics",uid="5b57d48f-e7d3-4050-9cd1-16aaaa052996"}'
+        values: '1x130'
+      # Absent for the first 70 minutes, then present. `_x` is "no sample"; the scrape did not
+      # exist yet.
+      - series: 'up{container="oauth2-proxy",endpoint="metrics",instance="10.0.1.213:44180",job="jackett-auth",namespace="jackett",pod="jackett-c5bfcbfb8-c75f9",service="jackett-auth-metrics"}'
+        values: '_x70 1x60'
+    promql_expr_test:
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="jackett"}'
+        eval_time: 65m
+        exp_samples:
+          - labels: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="jackett",severity="warning"}'
+            value: 1
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",namespace="jackett"}'
+        eval_time: 80m
+        exp_samples: []
+
+  # ---------------------------------------------------------------------------------------------
+  # 3. THE REGRESSION GUARD FOR `== 1` — the scenario that exposed a real bug on 2026-10-01.
+  #
+  #    A target is DISCOVERED but every scrape FAILS: `up` exists with value 0. This is not
+  #    hypothetical — it is the exact state podmonitor.yaml creates today, because the app
+  #    namespaces' default-deny drops the scrape (measured; see podmonitor.yaml).
+  #
+  #    With the original `unless count by (namespace) (up{container="oauth2-proxy"})`, the `up`
+  #    SERIES EXISTS, so `count by (namespace)` returns 1, `unless` removes the namespace, and the
+  #    alert GOES QUIET while nothing is being scraped at all — a green signal standing in for an
+  #    absent one, which is the precise disease this whole rule was written to cure. Deleting the
+  #    `== 1` makes THIS test fail and every other test in this file still pass, which is the
+  #    property that makes it worth having.
+  - interval: 1m
+    name: a discovered-but-FAILING target does not count as observed (up==0 must still fire)
+    input_series:
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://937ae6fc",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="jackett",pod="jackett-6877fcd9f5-ht9d9",service="kube-prometheus-stack-kube-state-metrics",uid="5b57d48f-e7d3-4050-9cd1-16aaaa052996"}'
+        values: '1x130'
+      # The PodMonitor target exists and never succeeds: NetworkPolicy drops the packets, so the
+      # scrape times out and `up` is published as 0 rather than going absent.
+      - series: 'up{container="oauth2-proxy",endpoint="metrics",instance="10.0.0.189:44180",job="oauth2-proxy-sidecars",namespace="jackett",pod="jackett-6877fcd9f5-ht9d9"}'
+        values: '0x130'
+    promql_expr_test:
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="jackett"}'
+        eval_time: 65m
+        exp_samples:
+          - labels: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="jackett",severity="warning"}'
+            value: 1
+      # And the inner clause itself yields NOTHING while the scrape is failing — which is the
+      # mechanism the alert depends on, asserted directly so a future reader can see why the
+      # `== 1` matters without re-deriving it. (Scenario 4 below covers the other direction:
+      # a succeeding scrape really does clear the alert, so `== 1` has not made the rule
+      # impossible to satisfy.)
+      - expr: 'count by (namespace) (up{container="oauth2-proxy"} == 1)'
+        eval_time: 65m
+        exp_samples: []
+
+  # ---------------------------------------------------------------------------------------------
+  # 4. A FLAPPING target: up goes 1 -> 0 and stays there. The alert must come BACK, not latch
+  #    cleared. Covers the ordering bug where a once-observed namespace is permanently excluded.
+  - interval: 1m
+    name: a target that starts working and then fails re-fires
+    input_series:
+      - series: 'kube_pod_container_info{container="oauth2-proxy",container_id="containerd://937ae6fc",endpoint="http",image="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",image_id="quay.io/oauth2-proxy/oauth2-proxy@sha256:d62e2d81",image_spec="quay.io/oauth2-proxy/oauth2-proxy:v7.8.1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="radarr",pod="radarr-7d55cbffb9-42hmh",service="kube-prometheus-stack-kube-state-metrics",uid="3333cccc-0000-0000-0000-000000000001"}'
+        values: '1x250'
+      # Healthy for the first 60m, then every scrape fails for the rest.
+      - series: 'up{container="oauth2-proxy",endpoint="metrics",instance="10.0.1.77:44180",job="oauth2-proxy-sidecars",namespace="radarr",pod="radarr-7d55cbffb9-42hmh"}'
+        values: '1x60 0x180'
+    promql_expr_test:
+      # Quiet while it was working.
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",namespace="radarr"}'
+        eval_time: 55m
+        exp_samples: []
+      # Firing again an hour after it broke.
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="radarr"}'
+        eval_time: 125m
+        exp_samples:
+          - labels: 'ALERTS{alertname="OAuth2ProxyGateUnobserved",alertstate="firing",namespace="radarr",severity="warning"}'
+            value: 1
+
+  # ---------------------------------------------------------------------------------------------
+  # 5. NO GATE, NO ALERT. A namespace with no oauth2-proxy container must never appear, or the
+  #    rule would page for every app on the platform the day someone widened the matcher.
+  - interval: 1m
+    name: namespaces without a gate sidecar are never reported
+    input_series:
+      - series: 'kube_pod_container_info{container="plex",container_id="containerd://feed0001",endpoint="http",image="ghcr.io/example/plex:1",image_id="ghcr.io/example/plex@sha256:feed0001",image_spec="ghcr.io/example/plex:1",instance="10.0.1.213:8080",job="kube-state-metrics",namespace="plex",pod="plex-0",service="kube-prometheus-stack-kube-state-metrics",uid="2222bbbb-0000-0000-0000-000000000001"}'
+        values: '1x130'
+    promql_expr_test:
+      - expr: 'ALERTS{alertname="OAuth2ProxyGateUnobserved"}'
+        eval_time: 65m
+        exp_samples: []
@@ -0,0 +1,137 @@
+---
+# oauth2-proxy AUTH-VISIBILITY alerting (R-OBS-04).
+#
+# ============================================================================================
+# WHY THIS FILE EXISTS — THE 2026-09-30 JACKETT FAILURE THAT NOTHING OBSERVED
+# ============================================================================================
+# Jackett's UI stopped working. Reloading never fixed it and the user was never re-prompted to
+# log in. The sidecar's own access log had the whole story:
+#
+#   [oauthproxy.go:1026] No valid authentication in request. Initiating login.
+#   GET "/api/v2.0/indexers/yts/config?_=..."          HTTP/1.1 ... 302 400
+#   GET "/api/v2.0/indexers/thepiratebay/config?_=..." HTTP/1.1 ... 302 409
+#   ... ten of these at 15:30:53, then a fresh [AuthSuccess] at 15:31:06,
+#   then "No valid authentication in request" again 26 SECONDS LATER, inside the same session.
+#
+# NOTHING ANYWHERE ELSE RECORDED IT. oauth2-proxy answers an unauthenticated XHR with a 302 to
+# Keycloak rather than a 401; the browser follows it silently, the live SSO session satisfies it
+# without a prompt, and the XHR resolves 200 carrying a 44 KB HTML dashboard where the app's
+# JavaScript wanted JSON. From the browser, from the user, and from any uptime check, the service
+# is healthy. `jackett-down` (apps/wave-0/jackett.yaml) watches
+# kube_deployment_status_replicas_available, which was 1 throughout and would have been 1 if the
+# gate had rejected every request ever made.
+#
+# ============================================================================================
+# WHY THIS FILE DOES NOT CONTAIN THE OBVIOUS ALERT — MEASURED, prod, 2026-09-30
+# ============================================================================================
+# The alert you would want is "a gated app is serving a sustained rate of auth redirects to a
+# user who is already authenticated". It is NOT WRITABLE against this cluster's TSDB, and the
+# reason is this file's real subject:
+#
+#   THE PER-APP GATE SIDECARS EXPORT NO METRICS AT ALL.
+#
+#   $ promtool query instant :9090 'count by (__name__) ({__name__=~"oauth2_proxy.*"})'
+#     oauth2_proxy_requests_in_flight => 2          <- the SHARED proxy's two replicas. That is ALL.
+#
+#   PARTLY SUPERSEDED 2026-10-01, and the correction matters more than the original. PR #28 gave
+#   the five FLEET sidecars `--metrics-address=0.0.0.0:44180` and a named `metrics` port, and the
+#   endpoint genuinely serves (verified by payload: http=200, 10208 bytes, all three
+#   `oauth2_proxy_*` families present — not merely by the flag being in `args`). podmonitor.yaml
+#   in this directory now discovers them. But the scrape is still DROPPED by the app namespaces'
+#   default-deny ingress, so the TSDB remains empty for those namespaces; and crafty still has no
+#   metrics listener at all. The rule below is unchanged in purpose and still has work to do.
+#
+#   Every sidecar (crafty, jackett, pihole, qbittorrent, radarr, sonarr) ran with
+#   `--http-address=0.0.0.0:4180` and NO `--metrics-address`, so oauth2-proxy's metrics listener
+#   was never started; the container declared only port `http:4180`, there was no `metrics` port on
+#   any of their Services, and there was no ServiceMonitor or PodMonitor for any of them anywhere
+#   in the cluster (measured: `kubectl get servicemonitors,podmonitors -A`). Six gates, zero
+#   telemetry. An alert written against `oauth2_proxy_requests_total{namespace="jackett"}` would
+#   have been PERMANENTLY INERT — present, evaluating, silent, and indistinguishable from health.
+#   This repo has shipped that mistake before (VeleroNoRecentBackup, the Longhorn rules), which is
+#   why `conformance/tests/l0-static/alert-metrics-have-scrape-targets` exists, and why the
+#   temptation was refused here rather than ratcheted into a register.
+#
+#   The SHARED proxy IS scraped (job="oauth2-proxy", 2 targets, up=1) but it is not in the data
+#   path for any sidecar-gated app — each app gates itself — and its current pods export no
+#   `oauth2_proxy_requests_total` at all, having proxied nothing since they started. An alert on
+#   the shared instance would not have fired on this incident either. Saying so is the point.
+#
+# So this file alerts on the CONDITION THAT MADE THE INCIDENT INVISIBLE, not on the incident.
+# That condition is measurable today, from two already-scraped metric families, and it is the
+# precondition for ever being able to alert on the incident itself. It self-clears the moment a
+# sidecar gains a metrics endpoint and a scrape — at which point the auth-churn rule becomes
+# writable and belongs here beside it. See runbooks/oauth2-proxy-auth-failures.md, which is the
+# by-hand equivalent an operator uses in the meantime.
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: oauth2-proxy-auth-visibility
+  namespace: oauth2-proxy
+  labels:
+    release: kube-prometheus-stack     # kube-prometheus-stack's ruleSelector
+    app.kubernetes.io/name: oauth2-proxy
+spec:
+  groups:
+    - name: oauth2-proxy.auth-visibility
+      rules:
+        - alert: OAuth2ProxyGateUnobserved
+          # A namespace runs an oauth2-proxy gate container, and Prometheus holds no scrape
+          # target for that container in that namespace.
+          #
+          # BOTH SIDES ARE ALREADY-SCRAPED FAMILIES, deliberately:
+          #   * kube_pod_container_info  — kube-state-metrics, the declared owner in
+          #     alert-metrics-have-scrape-targets. It sees the container whether or not the
+          #     container exports anything, which is exactly the property needed here.
+          #   * up                       — UNIVERSAL. `up` carries the `container` label for any
+          #     ServiceMonitor/PodMonitor target, so this works for a future per-app scrape
+          #     without knowing what job name it will pick.
+          # Nothing in this expression touches `oauth2_proxy_*`, so the rule cannot itself be
+          # inert for the reason it is reporting.
+          #
+          # MEASURED ON PROD 2026-09-30 — this expression returned six namespaces:
+          #   {namespace="crafty"} {namespace="jackett"} {namespace="pihole"}
+          #   {namespace="qbittorrent"} {namespace="radarr"} {namespace="sonarr"}
+          # and excluded namespace="oauth2-proxy", whose two replicas ARE scraped. It fires on
+          # the state the cluster is in; it is not a rule waiting for a hypothetical.
+          #
+          # `== 1` IS LOAD-BEARING, AND IT IS A BUG FIX (2026-10-01). This clause was originally
+          # a bare `up{container="oauth2-proxy"}`, which was WRONG in the one way that matters
+          # here: a target that is DISCOVERED BUT FAILING still publishes `up`, with value 0. So
+          # the series exists, `count by (namespace)` returns 1, `unless` removes the namespace,
+          # and the alert GOES QUIET while nothing whatsoever is being scraped. That is this
+          # branch's own disease — a green signal standing in for an absent one — reproduced
+          # inside the alert written to cure it. It was not hypothetical: podmonitor.yaml lands
+          # five targets whose scrape is DROPPED by NetworkPolicy (measured; see that file), so
+          # the bare form would have silently cleared on the very next commit. `== 1` means
+          # "observed" requires a scrape that actually succeeds.
+          #
+          # A failing target is therefore reported by BOTH this rule and kube-prometheus-stack's
+          # generic TargetDown. That duplication is deliberate: TargetDown says "a target is
+          # down", this says "an authentication gate is unobserved", and during an incident those
+          # are different sentences. The overlap is the cheap half of the trade.
+          expr: |
+            count by (namespace) (kube_pod_container_info{container="oauth2-proxy"})
+              unless
+            count by (namespace) (up{container="oauth2-proxy"} == 1)
+          # 1h, not 5m: a rollout briefly has a running sidecar whose target has not been
+          # discovered yet, and this condition is a standing architectural gap, never an
+          # emergency. Long enough that no deploy can produce it.
+          for: 1h
+          labels:
+            severity: warning
+          annotations:
+            summary: "oauth2-proxy gate in {{ $labels.namespace }} exports no metrics"
+            description: >-
+              The {{ $labels.namespace }} namespace runs an oauth2-proxy gate container and
+              Prometheus has NO scrape target for it, so every authentication decision that gate
+              makes — every redirect to Keycloak, every 403 on a group check, every session that
+              silently fails to hold — is invisible to metrics and to alerting. This is not a
+              theoretical gap: on 2026-09-30 the jackett gate answered a burst of the app's own
+              XHR calls with 302s to Keycloak, re-established a session, and lost it again 26
+              seconds later, and the only record anywhere was a line in the sidecar's stdout.
+              No alert could have fired, because there is nothing to alert on. Until a scrape
+              exists the runbook's LogQL queries are the substitute; the fix is
+              `--metrics-address=0.0.0.0:44180` on the sidecar plus a `metrics` port and a
+              ServiceMonitor, after which the auth-churn rule becomes writable.
+            runbook_url: "https://${BASE_DOMAIN}/runbooks/oauth2-proxy-auth-failures"
@@ -14,3 +14,12 @@ resources:
  - oauth2-proxy-credentials.sops.yaml
  - httproute.yaml       # forward-auth sign-in/callback ingress (was referenced in-place by platform-config)
  - servicemonitor.yaml  # Prometheus scrape (was referenced in-place by platform-config)
+  # Alerting on whether the gates are OBSERVABLE at all. Travels with the servicemonitor above
+  # for the same reason it does: monitoring.coreos.com CRDs are registered by the `platform`
+  # tier, which is Ready before this tier's platform-config dependency completes.
+  - auth-visibility-alerts.yaml
+  # Discovers the PER-APP gate sidecars' :44180 metrics endpoints by pod label (PR #28 added the
+  # listener; no Service exposes it, so this is a PodMonitor, not a ServiceMonitor). The scrape is
+  # still denied by the app namespaces' default-deny ingress — read podmonitor.yaml's header
+  # before concluding the targets should be up.
+  - podmonitor.yaml
@@ -0,0 +1,109 @@
+---
+# PodMonitor for the PER-APP oauth2-proxy gate sidecars (R-OBS-04).
+#
+# ============================================================================================
+# WHY A PodMonitor AND NOT A ServiceMonitor
+# ============================================================================================
+# PR #28 gave all five fleet sidecars `--metrics-address=0.0.0.0:44180` and a named `metrics`
+# container port. It did NOT add a Service port, and measured on prod 2026-10-01 none of the five
+# Services exposes 44180:
+#
+#   jackett      jackett:     [('port-9117', 9117), ('authproxy', 4180)]
+#   radarr       radarr:      [('port-7878', 7878), ('authproxy', 4180)]
+#   sonarr       sonarr:      [('port-8989', 8989), ('authproxy', 4180)]
+#   qbittorrent  qbittorrent: [('port-8080', 8080), ('authproxy', 4180)]
+#   pihole       pihole:      [('port-80',     80), ('authproxy', 4180)]
+#
+# A ServiceMonitor selects Service PORTS, so it would have nothing to select. A PodMonitor targets
+# pod ports directly and bypasses Services entirely — so this needs no change to the Services, and
+# therefore no change to `catalog/render.py` or the `apps/wave-*` renders and goldens.
+#
+# SELECTOR: the `homelab.keithjackson.dev/auth-mode: forward-auth` POD LABEL, not a namespace list.
+# Measured 2026-10-01: that label is present on exactly the five fleet gate pods (jackett, radarr,
+# sonarr, qbittorrent, pihole) and absent from crafty's, which is a separately-managed deployment
+# (platform/crafty) that still has no `--metrics-address` at all. So the selector describes the
+# PROPERTY being scraped rather than enumerating today's apps, and a sixth forward-auth app is
+# picked up with no edit here. `namespaceSelector.any: true` because those five namespaces are
+# created by the apps-wave-0/1 tiers, which this tier does not and must not depend on.
+#
+# ============================================================================================
+# READ THIS BEFORE BELIEVING THE TARGETS ARE UP: THE SCRAPE IS CURRENTLY DENIED BY NETWORKPOLICY
+# ============================================================================================
+# This PodMonitor is NECESSARY AND NOT SUFFICIENT. The endpoint serves, and the path from ns
+# monitoring is blocked. Both halves measured on prod 2026-10-01, read-only:
+#
+#   THE ENDPOINT IS HEALTHY — port-forward straight to the jackett pod's 44180:
+#     GET /metrics -> http=200 bytes=10208
+#     # TYPE oauth2_proxy_requests_in_flight gauge
+#     # TYPE oauth2_proxy_requests_total counter
+#     # TYPE oauth2_proxy_response_duration_seconds histogram
+#     oauth2_proxy_requests_total{code="200"} 3
+#     oauth2_proxy_requests_total{code="302"} 2
+#
+#   THE SCRAPE PATH IS DROPPED — from a pod in ns monitoring (grafana, which sits beside
+#   Prometheus and so occupies the same policy position) to the jackett pod IP:
+#     http://10.0.0.189:44180/metrics   -> wget: download timed out
+#     http://10.0.0.189:4180/ping       -> wget: download timed out     (even the "allowed" port)
+#     http://<shared-proxy>:44180/metrics -> 200, oauth2_proxy_* present (control: ns oauth2-proxy
+#                                            has NO NetworkPolicy, which is the only reason the
+#                                            shared instance has ever been scrapable)
+#
+#   WHY: ns jackett's `default-deny` NetworkPolicy is `podSelector: {}` with
+#   `policyTypes: [Ingress, Egress]` and an `egress` rule ONLY — no `ingress` key, which in
+#   NetworkPolicy semantics is deny-all-ingress. The only ingress allowance is the CNP
+#   `jackett-allow-gateway-ingress`: `fromEntities: [ingress]` to port 4180 — the Gateway's Envoy
+#   and nothing else. Prometheus matches no clause and its packets are dropped, not refused.
+#
+# SO THE TARGETS BELOW WILL READ `up == 0` UNTIL AN INGRESS ALLOWANCE LANDS. That is a known,
+# measured, documented state — NOT a regression introduced here, and NOT something to silence.
+# It is deliberately left visible: `OAuth2ProxyGateUnobserved` in auth-visibility-alerts.yaml
+# requires `up == 1`, precisely so a discovered-but-failing target does not read as "observed"
+# (see that file's header — this exact trap is why the `== 1` is there).
+#
+# THIS IS THE SAME BUG LONGHORN HAD, and the platform has a documented fix idiom for it:
+# platform/longhorn/networkpolicy-metrics.yaml, whose header records the identical fingerprint
+# measured on 2026-09-13 — ServiceMonitor in place, 3 targets DISCOVERED, `up` 0/0/0, TSDB empty,
+# `lastError: context deadline exceeded` (a drop, not a refusal). The fix there is a narrow
+# additive `networking.k8s.io/v1` NetworkPolicy: ns monitoring -> the pod label, that one port.
+#
+# WHY THAT POLICY IS NOT IN THIS DIRECTORY. It would have to be five objects, one per app
+# namespace, and those namespaces are created by apps-wave-0/1. Measured Flux graph, 2026-10-01:
+#   oauth2-proxy  dependsOn=['platform-config']        <- this tier
+#   apps-wave-0   dependsOn=['platform-config']        <- SIBLING, no ordering between them
+#   apps-wave-1   dependsOn=['apps-wave-0']
+# They are siblings, so on a cold bootstrap there is no guarantee ns jackett exists when this tier
+# applies, and a NetworkPolicy naming a missing namespace fails to apply. It would work on today's
+# running cluster and break on a rebuild — the "works now, broken on restore" shape this repo
+# keeps paying for. Longhorn's precedent is safe only because longhorn-system is created by the
+# same tier that ships its policy. The allowance therefore belongs WITH THE APPS, rendered next to
+# the rest of each app's network policy — i.e. in the catalog/renderer, which is in flight on
+# another branch. Tracked in the handoff; it is one narrow rule per app, not a redesign.
+apiVersion: monitoring.coreos.com/v1
+kind: PodMonitor
+metadata:
+  name: oauth2-proxy-sidecars
+  namespace: oauth2-proxy
+  labels:
+    release: kube-prometheus-stack     # kube-prometheus-stack's podMonitorSelector
+    app.kubernetes.io/name: oauth2-proxy
+    homelab.keithjackson.dev/purpose: metrics-scrape
+spec:
+  namespaceSelector:
+    # The gate sidecars live in the app namespaces, not this one. `any` rather than a fixed list so
+    # a new forward-auth app needs no edit here — the pod label below is what scopes this.
+    any: true
+  selector:
+    matchLabels:
+      # The PROPERTY, not the app list. Present on exactly the five fleet gate pods (measured
+      # 2026-10-01); absent from crafty, which has no metrics listener yet.
+      homelab.keithjackson.dev/auth-mode: forward-auth
+  podMetricsEndpoints:
+    # By PORT NAME (`metrics`, 44180), never by number: the name is what PR #28 declared on the
+    # container, and a numeric port here would silently stop matching if the port were renumbered.
+    - port: metrics
+      path: /metrics
+      interval: 30s
+      # 10s, matching the drop signature above: a DENIED path does not refuse, it hangs until the
+      # scrape times out. Leaving the default would make each failing target hold a scrape slot for
+      # the full interval.
+      scrapeTimeout: 10s
@@ -51,6 +51,36 @@ rule groups still evaluate against nothing and **cannot fire**:
 | Flux (`gotk_*`) | `FluxReconciliationFailure`, `FluxResourceSuspended`, `ImageUpdateAutomationFailing`, `ImagePolicyStalled`, `UpdateHealthGateFailed` |
 | Hubble (`hubble_drop_total`) | `HubbleDeniedFlowSpike` |
 | etcd (`kubeEtcd` disabled for k3s, deliberately) | no etcd fsync alerting exists at all |
+| per-app oauth2-proxy sidecars (`oauth2_proxy_*`) — updated 2026-10-01 | no authentication alerting exists at all for crafty, jackett, pihole, qbittorrent, radarr, sonarr |
+
+The oauth2-proxy row is the newest, and it has **moved twice in two days** — worth reading in order,
+because each step looked like the end of it.
+
+**2026-09-30 — nothing to scrape.** `count by (__name__) ({__name__=~"oauth2_proxy.*"})` returned
+exactly one family, `oauth2_proxy_requests_in_flight`, from the two replicas of the **shared**
+forward-auth instance. All six per-app gate sidecars ran without `--metrics-address`, so their metrics
+listener never started, and no ServiceMonitor or PodMonitor selected them.
+
+**2026-10-01 — the listener exists, and the scrape is still blocked.** PR #28 added
+`--metrics-address=0.0.0.0:44180` and a named `metrics` port to the five fleet sidecars, and the
+endpoint genuinely serves (port-forwarded: `http=200`, 10208 bytes, all three `oauth2_proxy_*`
+families, `oauth2_proxy_requests_total{code="302"}` present). `platform/oauth2-proxy/podmonitor.yaml`
+now discovers them by pod label. **But the app namespaces' `default-deny` NetworkPolicy drops the
+scrape**: from a pod in ns `monitoring`, both `:44180` and `:4180` on the jackett pod time out, while
+the shared proxy's `:44180` answers — the only difference being that ns `oauth2-proxy` has no
+NetworkPolicy at all. So the targets are discovered and `up == 0`, and the TSDB is still empty for
+those namespaces. **This is the same bug Longhorn had**, with the same fingerprint; the fix idiom is
+`platform/longhorn/networkpolicy-metrics.yaml` (ns monitoring → one pod label → one port), and for
+these apps it has to be rendered alongside each app's own policy rather than shipped from the
+oauth2-proxy tier, which is a *sibling* of `apps-wave-0` in the Flux graph and so cannot rely on those
+namespaces existing yet.
+
+`OAuth2ProxyGateUnobserved` (`platform/oauth2-proxy/auth-visibility-alerts.yaml`) alerts on the gap
+itself from `kube_pod_container_info` and `up` — two families that *are* scraped. Note its `up == 1`:
+a discovered-but-failing target publishes `up` with value **0**, so the earlier bare form would have
+gone quiet the moment the PodMonitor landed, while nothing was being scraped. That near-miss is the
+whole reason to distrust a cleared alert here. Crafty remains outside all of this — no metrics
+listener at all. See `oauth2-proxy-auth-failures`.

 Each affected runbook says so on its own page and gives the by-hand equivalent. See `targets-down`.

@@ -108,6 +138,14 @@ Each affected runbook says so on its own page and gives the by-hand equivalent.
 - **`keycloak-db`** — Keycloak's CloudNativePG database is unavailable.
 - **`keycloak-client-secrets`** — the client/URL reconciler failing or silent. The 2026-09-06 lockout that
  every green signal missed.
+- **`oauth2-proxy-auth-failures`** — a gated app "is up but does not work". Linked from
+  `OAuth2ProxyGateUnobserved`. The gate answers an unauthenticated **XHR** with a 302 to Keycloak
+  rather than a 401, the browser follows it silently against a live SSO session, and the call
+  resolves **200 carrying HTML** where the app's JavaScript wanted JSON — so the user is never
+  re-prompted and every uptime check stays green. Tells the gate's refusal from the group check
+  from the app's own login, from the sidecar log and from Loki, with the 2026-09-30 jackett trace
+  as the worked example. **Read the caveat**: the per-app gate sidecars export no metrics at all,
+  so these logs (72h of them) are the only record this class of failure leaves.

 ### Operator procedures (no alert links here)
 - **`plex-claim`** — claiming the Plex Media Server. **Claimed 2026-09-29**; the page now only
@@ -115,6 +153,18 @@ Each affected runbook says so on its own page and gives the by-hand equivalent.
  plex.tv/claim — so it is a timed sequence, scripted as `infra/plex/claim-ship.sh`:
  token → `sops` → PR → gate → merge → **forced** `flux reconcile` → pod roll (~2m45s measured).
  Includes why a changed Secret does not restart anything, and why an expired token is harmless.
+- **`plex-quicksync`** — hardware transcoding on the host iGPU. **Confirmed working 2026-10-02**
+  (VAAPI decode *and* encode, zero-copy). GitOps requests `gpu.intel.com/i915` on the pod
+  (`hardware.gpu: intel-quicksync`, satisfied by the `intel-gpu-plugin` DaemonSet — a `/dev/dri`
+  hostPath was measured not to work: device-cgroup `EPERM`); **a human still has to tick Settings →
+  Transcoder → "Use hardware acceleration when available"**, and no environment variable can do it
+  (the image supports no `PLEX_PREFERENCE*`, verified). **The proof is `Plex Media Server.log`'s
+  `final decoder: / final encoder:` line, not the Plex dashboard** — the dashboard showed nothing
+  while hardware transcoding was working, so its silence proves nothing; the page carries the
+  redacting grep (`X-Plex-Token` lives in that log), the WRONG log file to avoid, and two more
+  documented false leads (an empty `/proc/*/fd` scan, and a missing `HardwareAcceleratedCodecs` key).
+  Plex renders no ServiceMonitor, so no alert could ever fire. Also: **`Pending` stalls the whole app
+  tier** and its one-field escape hatch, and why a `:latest` pull can silently break it.
 - **`oauth2-proxy-session`** — a gated UI loads but its saves quietly do nothing, and you are never
  re-prompted to log in. Tells the three stacked auth layers apart from the sidecar's own access log
  (the four lines that matter, including the upstream `-` that means the proxy answered a request the
@@ -85,16 +85,29 @@ configMapGenerator:
      - node-io=node-io.md
      - node-memory=node-memory.md
      - node-memory-swap=node-memory-swap.md
+      # Linked from OAuth2ProxyGateUnobserved, and the diagnostic procedure for any gated app
+      # that "is up but does not work" — the gate answers an unauthenticated XHR with a 302,
+      # not a 401, so the failure arrives as an HTTP 200 carrying the wrong body.
+      - oauth2-proxy-auth-failures=oauth2-proxy-auth-failures.md
      # Operator procedure, not alert-linked. Reading the per-app oauth2-proxy sidecar log to tell
      # Keycloak, the sidecar and the app's own login apart. Cookie collision between two sidecars
      # presents as random session loss in ONE app's log and is only legible with both side by side,
      # so the page has to be reachable at 3am without going looking for it in git.
+      # SIBLING PAGE to oauth2-proxy-auth-failures above, and they are deliberately both here: this
+      # one is the SESSION-LIFECYCLE procedure (why a session died), that one is the DIAGNOSTIC
+      # TRIAGE (which of the three layers refused, and how to tell a 302-to-200 from a real success).
      - oauth2-proxy-session=oauth2-proxy-session.md
      # Operator procedure, not an alert page (no alert links to it — the
      # alert-runbooks-resolvable check treats that as advisory). It is served
      # because the plex.tv claim window is short and is not a thing to go looking
      # for in git: /runbooks/plex-claim has to resolve the moment it is needed.
      - plex-claim=plex-claim.md
+      # Operator procedure, not an alert page, and NO alert is possible for it: plex renders no
+      # metricsEndpoint, so there is no ServiceMonitor, no transcode-mode metric and nothing a rule
+      # could evaluate. Served because the feature it completes is INERT until a human ticks a box in
+      # the Plex UI — the one step GitOps cannot reach — and a page nobody can find is how a shipped
+      # feature stays switched off for months.
+      - plex-quicksync=plex-quicksync.md
      - pvc-capacity=pvc-capacity.md
      - targets-down=targets-down.md
      - update-rollback=update-rollback.md
@@ -0,0 +1,244 @@
+# oauth2-proxy auth failures — telling the three layers apart
+
+Linked from `OAuth2ProxyGateUnobserved`, and the page to open whenever a gated app "is up but
+does not work": a UI that loads and then shows nothing, a page that never finishes, a list that
+stays empty, a save that silently does not save. The alert is about the *visibility* gap; this
+page is the diagnostic procedure, and it works whether or not the alert ever fired.
+
+## What this means here
+
+Every app behind Keycloak on this platform has **three** places a request can be refused, and
+from a browser they look identical — because two of the three answer with **HTTP 200**.
+
+| Layer | What refuses | What the browser sees |
+|---|---|---|
+| 1. The gate (oauth2-proxy sidecar) | no valid session cookie | **302 to Keycloak**, then usually **200** |
+| 2. The gate's group check | session valid, wrong group | 403, or a redirect loop |
+| 3. The app itself | the app's own login / API key | the app's own 401 or login page |
+
+**Layer 1 is the one that hides.** oauth2-proxy answers an unauthenticated request with a
+redirect, not a 401 — correct for a browser navigation, disastrous for an **XHR**. The app's
+JavaScript asks for JSON, the gate answers `302`, the browser follows it silently, the live
+Keycloak SSO session satisfies it with no prompt, and the XHR resolves **200 with an HTML
+dashboard in the body**. `fetch()` sees a success. The user sees an empty page. Reloading does
+nothing, and nobody is ever asked to log in, because from the browser's point of view nothing
+failed.
+
+The worked example, jackett on 2026-09-30, straight from the sidecar log:
+
+    [oauthproxy.go:1026] No valid authentication in request. Initiating login.
+    GET "/api/v2.0/indexers/yts/config?_=..."          HTTP/1.1 ... 302 400
+    GET "/api/v2.0/indexers/thepiratebay/config?_=..." HTTP/1.1 ... 302 409
+    ... ten of these at 15:30:53
+    [AuthSuccess] Authenticated via OAuth2: Session{email:... created:2026-09-30 15:31:06
+                  expires:2026-09-30 15:36:06 refresh_token:true groups:[platform-admins]}
+    GET "/UI/Dashboard"        200 44379
+    GET "/api/v2.0/indexers?_=..." 200 1501229
+    ... and 26 seconds later, inside that same 5-minute session:
+    [oauthproxy.go:1026] No valid authentication in request. Initiating login.
+
+A session created at 15:31:06, valid until 15:36:06, not honoured at 15:31:32. That is the
+signature of the session being **re-established over and over instead of held** — a cookie that
+another proxy is overwriting, or a cookie that never comes back at all — not of Keycloak being
+down and not of the app being broken.
+
+**Read this before you trust any silence (updated 2026-10-01).** The five fleet gate sidecars now
+*do* run a metrics listener on `:44180`, and `platform/oauth2-proxy/podmonitor.yaml` discovers
+them — but **the scrape is dropped by the app namespaces' `default-deny` NetworkPolicy**, so those
+targets read `up == 0` and the TSDB holds nothing for them. Crafty has no listener at all. Only
+the shared `oauth2-proxy` namespace instance is genuinely scraped, and it is not in the data path
+for any of these apps. So there is still nothing to graph and nothing that could have alerted.
+**The logs below remain the only record that this class of failure happened.** Check which state
+you are in before you trust a quiet dashboard — "First commands" step 1 tells you in one query.
+
+## First commands
+
+`kubectl` does not exist on the workstation; everything runs in the pinned container, and always
+with an explicit context.
+
+**1. Which namespaces have an unobserved gate** (this is what the alert measures):
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n monitoring exec prometheus-kube-prometheus-stack-prometheus-0 \
+     -c prometheus -- /bin/promtool query instant http://localhost:9090 \
+     "count by (namespace) (kube_pod_container_info{container=\"oauth2-proxy\"}) unless count by (namespace) (up{container=\"oauth2-proxy\"} == 1)"'
+```
+
+**1b. If a namespace is listed, is it "no target" or "target failing"?** The two need different
+fixes and the alert deliberately does not distinguish them:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n monitoring exec prometheus-kube-prometheus-stack-prometheus-0 \
+     -c prometheus -- /bin/promtool query instant http://localhost:9090 \
+     "up{container=\"oauth2-proxy\"}"'
+```
+
+No series for the namespace → nothing discovers it (no PodMonitor match; check the pod carries
+`homelab.keithjackson.dev/auth-mode=forward-auth` and a named `metrics` port). A series with
+value **0** → discovered but every scrape fails; get the reason from the target's `lastError`:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n monitoring port-forward svc/kube-prometheus-stack-prometheus 19090:9090 & sleep 5; \
+   curl -s localhost:19090/api/v1/targets | python3 -c "
+import json,sys
+for t in json.load(sys.stdin)[\"data\"][\"activeTargets\"]:
+    if t[\"labels\"].get(\"container\")==\"oauth2-proxy\":
+        print(t[\"labels\"].get(\"namespace\"), t[\"health\"], t.get(\"lastError\",\"\")[:90])"'
+```
+
+`context deadline exceeded` is a **drop**, not a refusal — that is NetworkPolicy, not a dead
+endpoint. Confirm the endpoint itself is fine by port-forwarding straight to the pod, which
+bypasses policy entirely:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n jackett port-forward \
+     $(kubectl --context k3s-prod -n jackett get pod -l app.kubernetes.io/name=jackett -o name | head -1) \
+     14418:44180 & sleep 5; curl -s -o /dev/null -w "http=%{http_code} bytes=%{size_download}\n" \
+     http://127.0.0.1:14418/metrics'
+```
+
+`http=200` with a few KB means the listener is healthy and the path is what is broken. As of
+2026-10-01 that is the state of all five fleet apps, and the fix is an ingress allowance from ns
+`monitoring` to port 44180 — see `platform/longhorn/networkpolicy-metrics.yaml` for the idiom.
+
+**2. The sidecar log — but you must grep it.** The gate's stdout is ~11,500 lines a day per app
+and almost all of it is the kubelet health probe (`GET "/ping" ... 200`). A plain
+`kubectl logs --tail=100` shows you nothing but `/ping` and looks perfectly healthy. Filter:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n jackett logs deploy/jackett -c oauth2-proxy --since=2h \
+     | grep -v "/ping" | grep -E "No valid authentication|AuthSuccess|403|401|callback"'
+```
+
+Substitute the namespace: `crafty`, `jackett`, `pihole`, `qbittorrent`, `radarr`, `sonarr`.
+
+**3. The same question across every gate at once, from Loki.** Alloy collects all of these, so
+you can ask about the whole fleet and about a window that has already scrolled out of
+`kubectl logs`. Port-forward Loki and query it:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n monitoring port-forward svc/loki 13100:3100 & sleep 5; \
+   curl -s -G --data-urlencode "query=sum by (namespace) (count_over_time({container=\"oauth2-proxy\"} |= \"No valid authentication in request\" [24h]))" \
+     http://127.0.0.1:13100/loki/api/v1/query | python3 -m json.tool | grep -E "namespace|value" '
+```
+
+and the denominator, which is what makes the number mean something:
+
+```
+   ... --data-urlencode "query=sum by (namespace) (count_over_time({container=\"oauth2-proxy\"} |= \"AuthSuccess\" [24h]))" ...
+```
+
+**How to read the ratio.** A healthy sign-in is roughly **one or two** `No valid authentication`
+lines per `AuthSuccess` — the first request of a session is unauthenticated by definition. A
+ratio in the tens means sessions are not surviving. Measured on prod 2026-09-30 (24h):
+
+| namespace | `No valid authentication` | `AuthSuccess` | ratio |
+|---|---|---|---|
+| crafty | 81 | 1 | 81:1 |
+| radarr | 67 | 1 | 67:1 |
+| jackett | 39 | 4 | ~10:1 |
+| pihole, qbittorrent, sonarr | 0 | 0 | (unused in the window) |
+
+All three of those are the failure, not the baseline.
+
+**The cleanest worked case of all — radarr, measured 2026-10-01.** After the cookie fix landed,
+jackett dropped to **0** `No valid authentication` lines in 13,612 (from 39), and `removing
+session` / `cookie signature not valid` went to **0 on both apps** (radarr had 8 and 8 before).
+The eviction fingerprint is gone. But radarr still logged **206** rejections — and sampling 209 of
+them, **every single one was the same request**:
+
+    GET "/signalr/messages/negotiate" ... 302    <- browser UA, over and over
+
+That is one left-open browser tab. Its cookie is gone, its SignalR socket keeps trying to
+reconnect, and every attempt is silently 302'd to Keycloak. Nobody is at the keyboard, so nobody
+is ever re-prompted; the tab just sits there looking like a working page that has stopped
+updating. **This is the entire failure class in one line**: a non-navigation request answered with
+a redirect, forever, with no error surfaced anywhere.
+
+So when you read a rejection count, **always group by request path before concluding anything**:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n radarr logs deploy/radarr -c oauth2-proxy --since=24h \
+     | grep -v "/ping" | grep "No valid authentication" -A 1 \
+     | grep -oE "\"[^\"]+\" HTTP" | sort | uniq -c | sort -rn | head'
+```
+
+One path dominating means an orphaned client, not a broken gate — annoying but harmless, and it
+clears when the tab is closed. Many different paths from one session means the session itself is
+not surviving, which is the real defect. A per-app `--api-route` is the structural fix for the
+first shape: it makes those paths answer **401** instead of 302, so the client fails loudly
+instead of looping. Note that a count of **0 `AuthSuccess`** alongside a high rejection count
+means nobody has signed in since the last restart — so `--cookie-refresh` has never run, and no
+session-lifecycle conclusion can be drawn from that window at all.
+
+**4. Did the gate admit the user, or did the app refuse them?** Look for the `[AuthSuccess]` line
+and what follows it. If `[AuthSuccess]` appears and the very next thing is the *app's own* login
+page or the app's 401, the gate did its job and layer 3 is your problem — that is the Crafty
+shape, and `runbooks/crafty-admin-credential.md` covers it.
+
+**5. Cookie names, when sessions do not survive.** Two oauth2-proxy instances emitting the
+**same** cookie name on the **same** domain hold different cookie-secrets, so each overwrites the
+other's cookie and the loser logs exactly the `No valid authentication` line above. List what
+each gate is configured to emit (arguments only — never print a Secret's data):
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'for ns in crafty jackett pihole qbittorrent radarr sonarr; do printf "%-12s " "$ns"; \
+     kubectl --context k3s-prod -n $ns get deploy $ns \
+       -o jsonpath="{range .spec.template.spec.containers[?(@.name==\"oauth2-proxy\")]}{.args}{end}" \
+     | tr "," "\n" | grep -E "cookie-name|cookie-domain" | tr "\n" " "; echo; done'
+```
+
+A gate with no `--cookie-name` is using oauth2-proxy's default `_oauth2_proxy`. If two of them
+share that default **and** a zone-wide `--cookie-domain=.home.keithjackson.dev`, they are
+evicting each other, and the symptom is intermittent and depends on which app you visited last.
+
+## Likely causes, most likely first
+
+1. **Cookie collision between two gates on one domain.** Default cookie names plus a zone-wide
+   cookie domain. Signature: high `No valid authentication`-to-`AuthSuccess` ratio, sessions
+   dying seconds after being created, symptom depends on browsing order, and *nothing* wrong at
+   Keycloak. Fix is a unique `--cookie-name` and a host-scoped `--cookie-domain` per gate — see
+   `decisions/2026-09-29-crafty-oauth2-proxy-signin.md`, which found and fixed exactly this.
+2. **The cookie is too large to come back.** With `--pass-access-token` the session carries the
+   access token; a fat Keycloak token pushes the cookie past 4 KB and it gets split into
+   `_oauth2_proxy_0`, `_oauth2_proxy_1`… If a chunk is dropped the session is unreadable and the
+   gate logs the same line. Look for numbered cookie names in the log, and for a session that
+   never survives a single navigation.
+3. **The XHR path is not the navigation path.** The app's API is under a prefix the gate handles
+   differently, or the app issues cross-origin XHRs the cookie does not accompany. Signature:
+   `/UI/...` returns 200 and `/api/...` returns 302 in the *same second*, which is precisely the
+   jackett trace above.
+4. **Session lifetime shorter than the app's polling interval,** with no refresh. The jackett
+   session above was a 5-minute session. If `refresh_token:true` is present and sessions still
+   die early, the refresh is failing, not missing.
+5. **Group check, not authentication.** `--allowed-group` is a floor: a valid Keycloak session in
+   the wrong group is refused *after* a successful login, which reads as a loop. Look for 403 and
+   for the `groups:[...]` list on the `[AuthSuccess]` line.
+6. **The app, not the gate.** `[AuthSuccess]` followed by the app's own login page or 401.
+
+## When to escalate or stop
+
+- **Stop and hand over if the fix is a manifest change.** Everything above is read-only
+  diagnosis. Changing a `--cookie-name`, adding `--metrics-address`, or editing an alert goes
+  through a branch and a PR; `kubectl edit` on a gate is drift that the next Flux reconcile
+  reverts, and in the meantime you have an unreproducible cluster.
+- **Stop if you find yourself about to print a Secret.** The cookie-secret and client-secret live
+  in `oauth2-proxy-credentials` and the per-app equivalents. Compare *key names* and whether two
+  Secrets differ; never dump `.data` or `.stringData`. A password has been leaked into a
+  transcript here exactly this way.
+- **Escalate to the platform owner if Keycloak itself is the suspect** — check `KeycloakDown` and
+  `KeycloakDatabaseDown` first (`runbooks/keycloak-down.md`, `runbooks/keycloak-db.md`). In the
+  2026-09-30 incident Keycloak was healthy throughout and blaming it would have cost days.
+- **The honest limit of this page.** With no metrics from the sidecars there is no history beyond
+  Loki's **72h** retention (`platform/monitoring/loki.yaml`, `retention_period: 72h`). If the
+  report is older than three days, the evidence is gone and the only remaining move is to
+  reproduce it. That is the cost of the gap `OAuth2ProxyGateUnobserved` is tracking, and it is
+  the reason to close it rather than silence it.
@@ -0,0 +1,561 @@
+# plex-quicksync — making Plex actually use Intel Quick Sync, and proving it did
+
+**Status: hardware transcoding is CONFIRMED WORKING on prod as of 2026-10-02** — VAAPI for both
+decode *and* encode, zero-copy, measured in Plex's own server log (see "How to verify"). The manual
+toggle has been ticked. This page is therefore now mostly a **verification and regression**
+procedure, and its old verification section was **wrong** — see the correction at the top of
+"How to verify".
+**Alert:** none, and none is possible — see "How to verify" below. This is an **operator procedure**,
+not an incident page. It is indexed here because the feature it completes is **inert until a human
+clicks one button**, and a switched-off feature that looks shipped is exactly what this page exists
+to prevent.
+**Files:** `catalog/plex.yaml` (`hardware.gpu: intel-quicksync`), `apps/wave-0/plex.yaml` (the
+rendered golden — the `resources.limits: gpu.intel.com/i915: 1` block), `catalog/render.py` (the
+mechanism), `schemas/renderer-contract.md` `### hardware` (the contract),
+`platform/intel-gpu-plugin/` (the DaemonSet that advertises the resource).
+**Related:** `runbooks/plex-claim.md` — the other Plex procedure with a manual last mile.
+
+## STOP HERE IF PLEX IS `Pending` / `Unschedulable`
+
+That is this feature's **one serious failure mode** and it **stalls the whole app tier**
+(`apps-wave-0` is `wait: true`; `apps-wave-1` `dependsOn` it). Since 2026-10-01 Plex requests the
+extended resource `gpu.intel.com/i915: 1`, which is a **hard scheduling constraint** — if the
+`intel-gpu-plugin` DaemonSet is not advertising it on at least one schedulable node, Plex does not
+start at all. Go to **"Plex will not schedule"** below; it is the first troubleshooting section on
+purpose.
+
+**The escape hatch, in one sentence:** delete the two `hardware:` lines from `catalog/plex.yaml`,
+re-render, and merge — Plex returns to software transcoding and the tier unblocks. Exact commands
+are in **"Plex will not schedule" → remedy B**.
+
+## What this means here
+
+Plex transcodes in **software** by default. On the UHD Graphics 630 iGPU of an i7-9700T, Quick Sync
+moves that work off the CPU — which matters on a three-node cluster where the CPU is also running
+everything else. Two independent things have to be true before that happens, and **only one of them
+is in git**:
+
+| Half | Who does it | State |
+|---|---|---|
+| The render device is inside the container **and openable** | **GitOps.** `hardware.gpu: intel-quicksync` in `catalog/plex.yaml` renders `resources.limits: {gpu.intel.com/i915: 1}` on the app container. The `intel-gpu-plugin` DaemonSet (`platform/intel-gpu-plugin/`) injects the device **and the matching device-cgroup allow rule**. Flux applies it. | Done by merging that change. |
+| Plex is *told* to use it | **The operator, by hand, in the Plex web UI.** | **Done, 2026-10-02**, and confirmed from the **log** rather than from a preference key. `HardwareAcceleratedCodecs` is **still absent** from `Preferences.xml` with hardware transcoding working, so its absence proves nothing — see first-command #4. |
+
+**There is no GitOps path to the second half, and that is verified rather than assumed.** A bounded
+grep across the running image's entire init surface — `/etc/cont-init.d`, `/etc/services.d`,
+`/etc/fix-attrs.d`, `/plex-common.sh`, `/usr/local/bin` — found **zero** `PLEX_PREFERENCE*` support.
+Every `setPref` the image performs is one of six hardcoded keys (`MachineIdentifier`,
+`ProcessedMachineIdentifier`, `PlexOnlineToken`, `customConnections`, `allowedNetworks`,
+`TranscoderTempDirectory`), none of them the transcode toggle. No environment variable, no ConfigMap
+and no `postBuild` substitution can set it. So do not go looking for the clever way to automate this;
+there isn't one, and the search has already been done.
+
+### Why a resource request and not a `/dev/dri` hostPath
+
+Because the hostPath was **tried and measured not to work**, and knowing that is the only thing that
+will stop someone "simplifying" it back. Merged 2026-09-30, measured in the live pod on k3s-prod-2 on
+**2026-10-01**: the device node was plainly **visible** in the container
+(`crw-rw---- root video2 226,128 renderD128`) and **could not be opened** —
+
+| attempt | result |
+|---|---|
+| `uid 0`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+| `uid 1000`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+
+`EPERM` is the **device cgroup** refusing, not `EACCES` from file permissions. The two look identical
+from the outside and are not the same problem. Proven three ways: root on the **host** opened the
+same device `rc=0`; the node is unified `cgroup2fs`; and the container's runc `config.json` read
+`linux.resources.devices: [{allow: False, access: 'rwm'}]` with `linux.devices: None` — **one
+deny-all rule, no allow entry for major 226**. A hostPath is a *filesystem* operation: it binds the
+device node in and adds **no cgroup allow rule**, because Kubernetes cannot infer one from a volume.
+
+Two non-workarounds, so nobody spends an evening on them: **`uid 0` failed too**, so running as root
+does not bypass the cgroup gate; and **`privileged: true` is not available** — Kyverno
+`restrict-host-namespaces` is **Enforce** on app namespaces and rejects it.
+
+The device plugin is what produces the allow rule — in the v0.37.0 source,
+`createDeviceSpecsFromDrmFiles` emits `DeviceSpec{…, Permissions: "rw"}`, which the kubelet turns
+into both the device injection and the matching cgroup allow.
+
+**The failure mode that trade bought, stated plainly.** Before, a missing device meant the pod ran
+and **silently software-transcoded** — invisible, but harmless to everything else. Now, a missing
+*resource* means Plex is **Unschedulable** and the app tier stalls. That is strictly worse in blast
+radius and it is **unavoidable**: an extended resource is a hard scheduling constraint, and it is
+also the only thing that produces the cgroup allow rule. There is no graceful-degradation variant.
+The one improvement is that the loss is no longer silent.
+
+What this does **not** do, so nobody reads more into a green gate than is there: it does not enable
+hardware transcoding (that is the manual toggle below — done 2026-10-02), a green gate only proves
+the **manifest renders**, and it writes no `nodeSelector` — the resource itself constrains Plex to nodes the plugin
+advertises, which is the scheduler's doing and not a selector in this repo.
+
+## Running the commands in this runbook
+
+`kubectl` does not work on the operator workstation; the toolchain lives only in the pinned dev
+container. Every command below runs as:
+
+    infra/devcontainer/dev.sh run bash -c 'kubectl --context k3s-prod get nodes'
+
+Always pass `--context k3s-prod` explicitly. `dev.sh` merges both clusters into one kubeconfig and
+picks a current-context **per invocation**, and that default has already changed once. Do not run two
+`dev.sh` invocations at once.
+
+**Never `cat` `Preferences.xml`.** It contains `PlexOnlineToken`, a plex.tv account credential. Every
+read of that file below greps a single named key — do not widen one.
+
+## First commands
+
+**0. Is Plex running at all?** Ask this first now — the current mechanism can stop it starting.
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex -o wide'
+```
+
+`Pending` ⇒ go straight to **"Plex will not schedule"**. Running ⇒ carry on.
+
+**1. Do the nodes advertise the resource, and is the plugin Ready?** This replaces the old "is the
+hostPath mounted" check, and it is now the first thing that can be wrong.
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod get nodes -o json \
+   | jq -r ".items[] | \"\(.metadata.name) i915=\(.status.allocatable[\"gpu.intel.com/i915\"] // \"ABSENT\")\""'
+```
+
+Expect `i915=1` on each of the three prod nodes (measured 2026-10-01: all three). Any `ABSENT` ⇒ that
+node cannot run Plex. Then the DaemonSet behind it:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n intel-gpu-plugin get daemonset intel-gpu-plugin'
+```
+
+Expect **3 desired / 3 ready / 3 available** (measured 2026-10-01). Anything less is the cause, not a
+symptom — fix it there, in `platform/intel-gpu-plugin/`, not here.
+
+**2. Did Flux apply the current manifest?** `gpu.intel.com/i915` on the app container means the
+GitOps half landed:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex \
+     -o jsonpath="{.items[0].spec.containers[0].resources.limits}"; echo'
+```
+
+A `/dev/dri` **volumeMount** instead ⇒ the running pod predates the 2026-10-01 merge and is on the
+mechanism that does not work. Compare
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n flux-system get kustomization apps-wave-0 \
+     -o jsonpath="{.status.lastAppliedRevision}"; echo'
+```
+
+against the merge commit before blaming anything else.
+
+**3. Is the device visible inside the container, and can Plex *open* it?** This is the decisive test
+of the whole mechanism, and **it must be run as uid 1000**, because that is what Plex runs as.
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex exec deploy/plex -- ls -l /dev/dri'
+```
+
+Expect `renderD128`, mode `crw-rw----`. The plugin injects it; there is no longer any
+`DirectoryOrCreate` empty-directory case — if the resource were missing the pod would not be running
+to exec into. Then the open, **as the app user**:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex exec deploy/plex -- \
+     s6-setuidgid plex sh -c "exec 3<>/dev/dri/renderD128 && echo RW-OPEN-OK || echo RW-OPEN-DENIED"'
+```
+
+`RW-OPEN-OK` **as uid 1000** is the pass condition. **Passing as `uid 0` is weaker evidence and must
+not be accepted as proof** — Plex runs as 1000, and the whole point of the 2026-10-01 change was a
+cgroup rule that uid 0 could not buy its way past either. `RW-OPEN-DENIED` with the device plainly
+visible is the **device-cgroup** failure — i.e. the plugin's allow rule did not arrive; see "Likely
+causes" #2.
+
+**4. Is the toggle on?** **You cannot answer this from `Preferences.xml` alone.** That mistake is on
+this page's record: it told readers to treat a missing `HardwareAcceleratedCodecs` as "never
+enabled", and on 2026-10-02 that key was **absent from all 27 attributes** while hardware
+transcoding was demonstrably working. Named keys only — never the whole file, it holds
+`PlexOnlineToken`:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex exec deploy/plex -- \
+     grep -o -e "HardwareAcceleratedCodecs=\"[01]\"" -e "HardwareDevicePath=\"[^\"]*\"" \
+     "/config/Library/Application Support/Plex Media Server/Preferences.xml" || echo NEITHER-KEY'
+```
+
+| Reading | What it actually means |
+|---|---|
+| `HardwareDevicePath="…"` **present** | **Positive signal.** Plex writes this when hardware acceleration is configured; it was present on 2026-10-02 alongside a working VAAPI pipeline. |
+| `HardwareAcceleratedCodecs` **absent** | **Nothing. Neither on nor off.** Measured absent with the feature working (2026-10-02). Do not diagnose from it, and do not write it into the file to "fix" it. |
+| `HardwareAcceleratedCodecs="0"` | explicitly **off** — the only conclusive reading, and only in the negative. |
+| `HardwareAcceleratedCodecs="1"` | on. |
+
+**The log settles it, this file does not.** Go to "How to verify"; the `hardware transcoding:
+enabled` and `final decoder/final encoder` lines are the verdict.
+
+## The one manual step
+
+Plex web UI → **Settings → Transcoder → "Use hardware acceleration when available"** → tick → Save.
+(It sits under *Show Advanced*.) A transcode already in flight is unaffected; the next one picks it up.
+
+The underlying preference key is **`HardwareAcceleratedCodecs`**. That name was confirmed by
+extracting strings from the 23 MB `Plex Media Server` binary in the running container — **not** from
+recall and not from a forum post. **But do not expect to see it in `Preferences.xml` after ticking
+the box:** on 2026-10-02, with the toggle on and hardware transcoding working, the key was absent and
+`HardwareDevicePath` was what had been written. The key exists in the binary; Plex does not
+necessarily persist it. Real neighbouring keys from the same binary, so you can recognise them and
+not invent them:
+
+| Key | What it is |
+|---|---|
+| `HardwareAcceleratedCodecs` | the toggle this page is about |
+| `HardwareAcceleratedEncoders` | hardware **encode**, as distinct from decode |
+| `HardwareDevicePath` | how Plex is pointed at a *specific* render node — relevant only if a host ever exposes more than one |
+| `TranscoderToneMapping`, `TranscoderToneMappingAgorithm` | HDR→SDR tone mapping. **That misspelling of "Algorithm" is upstream's, in the binary. It is not a typo in this page — do not "fix" it** here or anywhere else; the corrected spelling is not a key Plex reads. |
+| `TranscoderHEVCEncoding` | HEVC output |
+
+**`HardwareAcceleratedCodecsEnableHEVC` does not exist.** It is widely cited and it is not in the
+binary. Recorded explicitly so nobody spends an evening hunting for it or, worse, writes it into
+`Preferences.xml` and concludes hardware transcoding is broken.
+
+Editing `Preferences.xml` by hand is **not** the procedure. It is app state on the replicated Longhorn
+tier, it is in no git commit, and the UI writes the key correctly. Use the UI.
+
+## How to verify — the log, not the dashboard
+
+**Correction, 2026-10-02.** Until today this page said the Plex dashboard showing `Transcode (hw)`
+was "the *only* proof available, and nothing else will ever tell you". **That was false and it cost
+the operator real time:** hardware transcoding was fully working — VAAPI decode *and* encode, with
+zero-copy — while **Dashboard → Activity showed nothing of the sort**. A check that is silent while
+the feature works does not just fail to prove anything; it sends someone hunting a fault that does
+not exist. The authoritative check is **Plex's own server log**.
+
+### The primary check — run this one
+
+```
+./infra/devcontainer/dev.sh run bash -c 'kubectl --context k3s-prod -n plex exec deploy/plex -- \
+  grep -aiE "hardware transcoding: (final|using)" \
+  "/config/Library/Application Support/Plex Media Server/Logs/Plex Media Server.log" \
+  | sed -E "s/(X-Plex-Token|token)=[^&[:space:]]*/\1=REDACTED/g" | tail -5'
+```
+
+| Line | Verdict |
+|---|---|
+| `hardware transcoding: final decoder: vaapi, final encoder: vaapi` | **PASS.** Hardware in **both** directions. This is the pass condition. |
+| `hardware transcoding: using zero-copy transcoding` | Best case — frames never round-trip through the CPU. Observed on prod 2026-10-02. |
+| `hardware transcoding: final decoder: , final encoder: ` (both empty) | It **fell back to software** for that session. |
+| no matching line at all | No transcode has been attempted since the log rolled — force one (below), then look again. Absence is not a failure. |
+
+The 2026-10-02 reading on prod, in full, for the shape to recognise:
+
+```
+Codecs: hardware transcoding: testing API vaapi for device '/dev/dri/renderD128'
+                              (Intel CoffeeLake-S GT2 [UHD Graphics 630])
+hardware transcoding: using hardware decode accelerator vaapi
+hardware transcoding: zero-copy support present
+hardware transcoding: using zero-copy transcoding
+hardware transcoding: final decoder: vaapi, final encoder: vaapi
+```
+
+To make a line appear, start a playback that **forces a transcode**: a client that cannot
+direct-play the source, or an explicitly lower quality. Direct Play / Direct Stream never reaches
+the transcoder and logs nothing.
+
+### Three things to get right about that command
+
+1. **Always keep the `sed`.** `Plex Media Server.log` carries API URLs containing
+   **`X-Plex-Token=…`**, a live plex.tv credential. Any output that might land in a ticket, a
+   transcript or a chat window must be redacted — CLAUDE.md §6, and a credential has already been
+   leaked into a transcript in this repo exactly this way. The redacting form above is the form to
+   copy; do not "simplify" the pipe out of it.
+2. **`Plex Transcoder Statistics.log` is the WRONG FILE.** The name is irresistible and someone will
+   reach for it. It holds XML `<Variant .../>` session records — **not** ffmpeg/VAAPI output. There
+   is no decoder/encoder verdict anywhere in it. The file you want is `Plex Media Server.log`.
+3. `grep -a` is deliberate: the log carries occasional non-UTF-8 bytes, and without `-a` GNU grep
+   reports "binary file matches" and prints nothing — which reads exactly like "no hardware
+   transcoding".
+
+### Secondary and unreliable: the Plex dashboard
+
+**Dashboard → Activity** *may* label a transcoding session `Transcode (hw)`. If it does, it agrees
+with the log and you are done. **If it does not, that proves nothing at all** — measured 2026-10-02:
+hardware transcoding active and zero-copy, dashboard silent. Never conclude "software transcoding"
+from the dashboard and never open an investigation on its say-so. Kept here only because it is the
+fastest glance available; it is a convenience, **not evidence**.
+
+### Process-level, while a transcode is in flight: the transcoder's argv
+
+This shows Plex *requesting* hardware, as distinct from reporting it afterwards. Get a shell in the
+pod (if `-it` has no TTY under `dev.sh run`, wrap each line in `sh -c "…"` instead):
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex exec -it deploy/plex -- sh'
+```
+
+then, inside the pod:
+
+```
+ps -eo pid,user,comm | grep -i transcoder
+tr '\0' ' ' < /proc/<transcoder-pid>/cmdline | grep -oE '(-hwaccel[a-z_]*|vaapi|renderD[0-9]+)'
+```
+
+`-hwaccel`, `-hwaccel_device`, `vaapi`, `renderD128` present ⇒ Plex is asking for hardware. Keep the
+`grep -oE`: the raw argv contains media paths and session identifiers you do not want in a
+transcript. This only works while a session is actually transcoding.
+
+### Documented false lead: scanning `/proc/*/fd` for a holder of `renderD128`
+
+**It returns nothing even when hardware transcoding is working.** Measured 2026-10-02: a
+`/proc/*/fd` sweep inside the pod found **no** process holding `/dev/dri/renderD128`, while the log
+showed `final decoder: vaapi, final encoder: vaapi` for sessions either side of the scan. Transcode
+sessions are **short** and the scan landed in a gap between two of them. An empty fd scan is a
+**timing artefact, not a finding** — do not report it as evidence of software transcoding, and do
+not escalate on it. Use the argv check above (which has the same timing constraint but tells you
+something when it does catch a session), or just read the log, which is retrospective and therefore
+does not care about timing.
+
+### Why there is no alert — this half was always right
+
+`catalog/plex.yaml` declares no `metricsEndpoint`, so Plex renders **no ServiceMonitor at all** —
+there is no scrape, no transcode-mode metric, and therefore no alert that could fire. Do not open
+work to "alert on software-transcode fallback": with no metric behind it such a rule would be
+permanently inert, and this repo has already been burned at scale by rules that evaluated against
+nothing while looking like coverage (`runbooks/README.md`, "Alerts you should not trust the silence
+of"). If that gap is ever worth closing, the honest order is a real exporter first and a rule
+second.
+
+## Likely causes, most likely first
+
+When the **log** says it fell back to software — `final decoder: , final encoder: ` with both
+fields empty. (Not when the dashboard is quiet: the dashboard's silence means nothing, see "How to
+verify".)
+
+1. **The toggle is off.** Historically the most likely answer and the reason this page exists. It was
+   ticked on 2026-10-02 and should stay ticked, but a `/config` restore or a stray UI save can unset
+   it. Remedy: the manual step above. **Do not diagnose this from a missing
+   `HardwareAcceleratedCodecs`** — see first-command #4; only `="0"` is conclusive, and the log line
+   `hardware transcoding: enabled` is the real signal that the toggle took.
+
+2. **The device cannot be `open()`ed — the plugin's cgroup allow rule did not arrive.** Symptom:
+   first-command #3 shows `renderD128` present and correctly owned, and the open returns
+   `RW-OPEN-DENIED` / `Operation not permitted`. This is `EPERM` from the **device cgroup**, not
+   `EACCES` from file permissions. **This is no longer a design risk — it is the measured reason the
+   design is what it is** (2026-10-01; see "Why a resource request and not a `/dev/dri` hostPath").
+   If it recurs *under the current mechanism*, the plugin is the suspect, not the manifest: confirm
+   the pod really did consume the resource
+
+   ```
+   infra/devcontainer/dev.sh run bash -c \
+     'kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex \
+        -o jsonpath="{.items[0].spec.containers[0].resources}"; echo'
+   ```
+
+   and then check the plugin pod on Plex's node (first-command #1). **Do not reach for
+   `privileged: true`** — Kyverno `restrict-host-namespaces` is Enforce and rejects it, and uid 0
+   failed the same open in 2026-09-30's measurement anyway. **Do not re-add a `/dev/dri` hostPath**;
+   that is the mechanism this replaced, and it was measured not to work.
+
+3. **The pod landed on a node whose GPU went away.** Under the current mechanism this does **not**
+   present as an empty `/dev/dri` — the node simply stops advertising `gpu.intel.com/i915` and Plex
+   either never schedules there or gets evicted off. So this cause now shows up as **`Pending`**, not
+   as silent software transcoding: see "Plex will not schedule". All three prod nodes carried
+   `renderD128` with `i915` bound when measured (2026-09-30) and all three advertised the resource on
+   2026-10-01, so this means a node changed — a rebuild, a BIOS reset, or a driver regression.
+   **Fix the node.**
+
+4. **The image changed under you.** See "The floating-image caveat". A `:latest` pull can remove the
+   mechanism this feature leans on with no commit in this repo.
+
+5. **The session was never transcoding.** Direct Play / Direct Stream never reaches the transcoder,
+   so no `hardware transcoding:` line is written for it at all. An *absent* line is this, not a
+   failure — confirm a transcode actually happened before treating anything as a hardware fault.
+
+6. **The *source* codec may be one this iGPU cannot decode — HYPOTHESIS, not a confirmed
+   diagnosis.** The symptom that points here is fallback on **particular titles** while other
+   sessions are hardware. Measured on prod 2026-10-02, *between* two fully hardware sessions, one
+   session logged:
+
+   ```
+   hardware transcoding: enabled, but no hardware decode accelerator found
+   hardware transcoding: final decoder: , final encoder:
+   ```
+
+   The platform was demonstrably working either side of it, so the likely explanation is the **file**,
+   not the cluster: UHD 630's media engine decodes H.264, HEVC 8-bit and 10-bit, and VP9 — but **not
+   AV1**, and not 10-bit 4:2:2 HEVC. **This was NOT verified against that particular title.** It is
+   inferred from the hardware's capability set and is written here as the lead to check first, not as
+   a finding. So: fallback on *all* files is a platform problem (causes 1-4); fallback on *some*
+   files, check the source codec before suspecting the device, and expect occasional legitimate
+   per-title fallbacks to be normal rather than a fault.
+
+## Plex will not schedule
+
+**Symptom:** the Plex pod sits `Pending`, and `apps-wave-0` does not go Ready, which holds
+`apps-wave-1` behind it. This is the cost of the 2026-10-01 mechanism and the section you want at
+2am.
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex describe pod -l app.kubernetes.io/name=plex | tail -20'
+```
+
+`0/3 nodes are available: 3 Insufficient gpu.intel.com/i915` (or similar) confirms it: **nothing is
+advertising the resource Plex requires**. Plex's schedulability is coupled to the `intel-gpu-plugin`
+DaemonSet, deliberately and unavoidably — the plugin is also the only thing that produces the
+device-cgroup allow rule, so there is no variant of this feature that degrades gracefully.
+
+**Remedy A — fix the plugin (preferred).** Run first-command #1. If the DaemonSet is not 3/3, or the
+nodes show `i915=ABSENT`, the fault is in `platform/intel-gpu-plugin/` or on the node, and that is
+where to fix it. Check its Flux Kustomization reconciled:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n flux-system get kustomization intel-gpu-plugin'
+```
+
+A plugin that is merely **not yet reconciled** resolves itself; one that is failing does not.
+
+**Remedy B — the escape hatch: give up hardware transcoding and unblock the tier.** One field, no
+code change. In a branch (never on `main`, never `kubectl edit` — Flux reconciles `main` and would
+revert you):
+
+1. delete these two lines from `catalog/plex.yaml`:
+
+   ```yaml
+   hardware:
+     gpu: intel-quicksync
+   ```
+
+2. re-render the golden:
+
+   ```
+   python3 catalog/render.py --all catalog -o apps --wave-layout --no-image-automation
+   ```
+
+   The only change must be `apps/wave-0/plex.yaml`, losing its `resources` block.
+3. gate, PR, merge. Plex schedules anywhere and **software-transcodes**; the app tier unblocks.
+
+Re-adding the field later is the same edit in reverse. Nothing about the Plex-side toggle or the
+`/config` PVC is affected either way — do **not** delete the PVC as part of this.
+
+## The floating-image caveat
+
+Everything this page states about the image's behaviour — the `cont-init.d` hook that adds the app
+user to the device's group, the absence of `PLEX_PREFERENCE*`, the six `setPref` keys, the
+`HardwareAcceleratedCodecs` string — was read from digest **`sha256:e0ab27395614…`**
+(`plexmediaserver 1.43.4.10903`).
+
+The deployment runs **`plexinc/pms-docker:latest` with `imagePullPolicy: Always`**. A silent pull can
+therefore change the behaviour this feature depends on **with no commit in this repo**, while the
+manifest still looks identical. Symptom to watch for: hardware transcoding that worked and quietly
+stopped, with no git change to point at.
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex get pod -l app.kubernetes.io/name=plex \
+     -o jsonpath="{.items[0].status.containerStatuses[0].imageID}"; echo'
+```
+
+If that digest is not the one above, **re-verify the hook before trusting this page**:
+
+```
+infra/devcontainer/dev.sh run bash -c \
+  'kubectl --context k3s-prod -n plex exec deploy/plex -- ls /etc/cont-init.d/'
+```
+
+`45-plex-hw-transcode-and-connected-tuner` still present ⇒ the group mechanism is still the one
+described here: the hook reads the injected device's **own** ownership, which under the device plugin
+is `root:root`, so it resolves GID `0` (see "When to escalate or stop"). Gone or renamed ⇒ stop and
+re-measure. In particular the "no `supplementalGroups` needed" conclusion in `catalog/plex.yaml` and
+`schemas/renderer-contract.md` is only true while that hook exists.
+
+## When to escalate or stop
+
+- **Do not add `supplementalGroups` or a `podSecurityContext` to `catalog/plex.yaml`.** It would be
+  **inert**, and it would mislead the next person. The image's `cont-init.d` hook stats each
+  `/dev/dri` char device and `usermod -a -G`s the `plex` user into **that device's** group, inside
+  the container; the service then execs via `s6-setuidgid plex`, which **rebuilds** the process group
+  list from `/etc/group` rather than appending to it. Anything Kubernetes supplies is therefore
+  discarded before Plex starts. `render.py --selftest` asserts that absence deliberately.
+- **GID `992` is a hostPath-era number — do not expect it, and do not read the root group as a
+  defect.** Under the device plugin the injected `/dev/dri/renderD128` arrives **`root:root`**, so
+  the hook resolves GID **`0`** and adds `plex` to the **root group**. Measured 2026-10-02:
+  `Plex Media Server` runs with `Groups: 0 100 1000`, and `id plex` reports
+  `groups=1000(plex),0(root),100(users)`. Under the old hostPath the device was `root:render`
+  (GID 992 on these nodes) and the reading was `Groups: 44 100 992 1000`. **992 is now irrelevant**:
+  if any page, comment or older handoff tells you to expect it, that text is stale and the system
+  you are looking at is working. The root-group membership is an inherent property of the plugin's
+  `DeviceSpec` — it injects the host device with its host ownership — **not** a defect, and **not** a
+  reason to reintroduce `supplementalGroups`, which would still be a no-op for the reason in the
+  bullet above.
+- **Do not reach for `privileged: true`.** Kyverno `restrict-host-namespaces` is Enforce and will
+  reject it — and it would not help: `uid 0` failed the same `open()` in the 2026-09-30 measurement,
+  because root does not bypass the cgroup device gate.
+- **Do not re-add a `/dev/dri` hostPath, and do not "simplify" the resource request into one.** That
+  was the original mechanism and it was measured non-functional on 2026-10-01 (device-cgroup `EPERM`
+  with the device plainly visible). The device plugin's `DeviceSpec{… Permissions: "rw"}` is what
+  grants access; a volume cannot.
+- **Do not `kubectl edit`/`patch` the Deployment to add or change the device mount.** Flux reconciles
+  `main`; a hand-applied change is drift that the next reconcile reverts or, worse, adopts. Mutating
+  `kubectl` and `flux reconcile` are **operator** actions, and the durable change is a git change
+  through a branch, a PR and a green gate.
+- **Do not edit `Preferences.xml` on the PVC**, never `cat` it, and never `sops -d` anything into a
+  terminal. It holds `PlexOnlineToken`.
+- **Do not delete the `plex-config` PVC to "start clean".** It is the library database, metadata and
+  watch state (`policy: daily-offcluster`). The transcode toggle is one preference, not a reason to
+  lose the DB.
+- **Stop and get the operator** if the device cannot be opened from the pod as uid 1000 (cause #2
+  confirmed) *under the current mechanism* — note that this open **passed** on 2026-10-02
+  (`RW-OPEN-OK` as uid 1000 via `s6-setuidgid plex`), so a failure now is a regression. The next step is a platform change in
+  `platform/intel-gpu-plugin/`, with its own PR, not more poking at this workload.
+
+## Known gaps, deliberately not fixed here
+
+Recorded so they are findable, and so nobody assumes they were overlooked:
+
+- **Transcode scratch is on the wrong medium, and it is invisible.** The image sets
+  `TranscoderTempDirectory=/transcode`, which is **not a volume** — it is the container's writable
+  overlay, i.e. the node's `/data`. That is NVMe on prod-2 but **SATA on prod-1 and prod-3**, so
+  transcode scratch performance changes silently whenever the pod is rescheduled. Both candidate fixes
+  need a **schema** change and are therefore out of scope here: pin Plex to prod-2 (needs a
+  `nodeSelector` field the schema does not have), or give `/transcode` a `medium: Memory` emptyDir and
+  transcode in RAM (~31 GiB per node; needs an emptyDir field the schema does not have). The RAM option
+  is the better one, because it removes the node asymmetry outright instead of working around it.
+- **GuC/HuC firmware is not loaded** (`i915.enable_guc=0`). The cost is low-power H.264 **encode**
+  quality only — decode, and the feature on this page, are unaffected. Changing it needs a kernel
+  parameter and a **serialized rolling reboot**, which is an operator action (CLAUDE.md §5), and this
+  cluster has a documented history of that going badly (a prod node hung on a warm reboot on
+  2026-09-03). Not worth coupling to this change.
+
+---
+Source of truth: `runbooks/plex-quicksync.md`. Index: `/runbooks/`
+
+**Hardware transcoding is CONFIRMED WORKING on prod, measured 2026-10-02.** From
+`Plex Media Server.log` in the running pod: API `vaapi` on `/dev/dri/renderD128`
+(`Intel CoffeeLake-S GT2 [UHD Graphics 630]`), `using hardware decode accelerator vaapi`,
+`zero-copy support present`, `using zero-copy transcoding`, and
+`final decoder: vaapi, final encoder: vaapi`. Also measured in the same pod: `RW-OPEN-OK` opening
+`/dev/dri/renderD128` as uid 1000 via `s6-setuidgid plex`; the pod spec carrying
+`{"limits":{"gpu.intel.com/i915":"1"},"requests":{"gpu.intel.com/i915":"1"}}`; `Groups: 0 100 1000`
+and `id plex` → `groups=1000(plex),0(root),100(users)`; and `HardwareDevicePath` present in
+`Preferences.xml` while `HardwareAcceleratedCodecs` was absent from all 27 attributes.
+
+Earlier readings, kept for provenance and **superseded** where they conflict: the node measurements
+(all three prod nodes) and the in-container image measurements taken **2026-09-30** against
+`plexinc/pms-docker` digest `sha256:e0ab27395614…` (`plexmediaserver 1.43.4.10903`); the hostPath
+`EPERM` open results and the `Groups: 44 100 992 1000` / GID-992 group reading, **2026-10-01, under
+the hostPath mechanism that no longer exists**; and the `gpu.intel.com/i915: 1` allocatable and
+`intel-gpu-plugin` DaemonSet readings, **2026-10-01**.
+
+What this page still does **not** establish: that any particular title's software fallback was caused
+by its source codec (cause 6 is an unverified hypothesis), and that Plex's dashboard *ever* labels a
+session `Transcode (hw)` on this deployment — it has not been observed doing so, which is exactly why
+the log is now the primary check.
@@ -0,0 +1,148 @@
+# plex-remote-access — getting Plex reachable from outside the LAN, and proving it
+
+**Status: the in-cluster half is shipped; the last mile is TWO MANUAL STEPS and remote access does
+NOT work until both are done.** Neither can live in this repo.
+**Alert:** none, and none is possible from in-cluster signals — see "Why no alert" below. This is an
+**operator procedure**, not an incident page.
+**Files:** `apps/wave-0/plex-remote.yaml` (the tcp/32400 LoadBalancer + its CiliumNetworkPolicy),
+`clusters/base/cluster-vars.yaml` + `clusters/prod/kustomization.yaml` (`PLEX_L4_VIP`),
+`catalog/render.py` (the `EXTRA_UNITS` entry that places the unit), `catalog/plex.yaml` (Plex's own
+env — note `ADVERTISE_IP` is deliberately NOT set; see "Why not the Gateway URL").
+**Related:** `runbooks/plex-quicksync.md` and `runbooks/plex-claim.md` — the other two Plex
+procedures with a manual last mile.
+
+## What was actually broken (measured 2026-10-03)
+
+Plex was serving correctly and was reachable on the LAN. Two independent things stopped remote
+clients:
+
+1. **Plex advertised no remote connection at all.** In the live `Preferences.xml`:
+   `customConnections` **ABSENT**, `ManualPortMappingMode` **ABSENT**, while
+   `PublishServerOnPlexOnlineKey="1"`. The server was trying to publish itself to plex.tv and had
+   nothing reachable to publish. Plex's automatic path (UPnP/NAT-PMP) cannot work here: the process
+   lives in a pod on the Cilium overlay (`10.0.2.136`), it cannot speak UPnP to the SOHO router, and
+   a mapping to the pod IP would be meaningless if it could.
+2. **Nothing from the WAN reached the cluster.** `136.51.115.252:32400` → connection refused.
+   `136.51.115.252:443` → a **self-signed router certificate** (`C=ZZ, CN=%D`), not the Gateway's
+   Let's Encrypt wildcard.
+
+The Service was `ClusterIP` only, so port 32400 had no path off-cluster by any route.
+
+**Caveat on that second measurement, because it is easy to over-read:** it was taken from *inside*
+the LAN, so strictly it proves there is no NAT hairpin (reflection). It does not prove outside-in
+failure. But a router certificate answering 443 means the router's own management UI is bound to the
+WAN port, and that *would* shadow a port-forward for genuine outside traffic too. **Verify from a
+real outside vantage point** (phone on cellular, or an external port checker) before concluding
+either way — this repo has no such vantage point and must not pretend otherwise.
+
+## Step 1 — forward WAN tcp/32400 on the router (operator, off-repo)
+
+    WAN tcp/32400  ->  192.168.2.24  (${PLEX_L4_VIP})  port 32400
+
+`192.168.2.24` is a pinned Cilium LB VIP, ARP-announced on the flat LAN by the existing
+`CiliumL2AnnouncementPolicy default-l2`. It is **pinned on purpose**: this address is typed into the
+router by hand, so it must never move on a reschedule.
+
+While in the router, **check what else is bound to the WAN interface.** If remote management is
+listening on 443 (see the certificate measurement above), move it or disable it — otherwise the
+Gateway's 80/443 forward is shadowed and the Plex *web UI* over the internet stays broken even
+though remote *streaming* works. These are two separate paths and they fail independently.
+
+Confirm the VIP is actually answering on the LAN first, so a failure later is unambiguous:
+
+    ./infra/devcontainer/dev.sh run bash -c \
+      'kubectl --context k3s-prod -n plex get svc plex-l4-tcp -o wide'
+
+`EXTERNAL-IP` must read `192.168.2.24`. If it is `<pending>`, LB-IPAM could not assign — the address
+is outside the pool or already taken; see `clusters/prod/kustomization.yaml` for the pool accounting.
+
+**`192.168.2.24` was never ARP-probed off-cluster.** Only the in-cluster pool accounting was checked
+(IPsTotal 11 / IPsUsed 4 / IPsAvailable 7 on 2026-10-03). If the VIP behaves oddly, probe it from a
+LAN host — `arping 192.168.2.24` — the way `.23` was probed before it was claimed.
+
+## Step 2 — tick "manually specify public port" in Plex (operator, UI only)
+
+**There is no GitOps path to this and no environment variable for it.** Measured in the running
+image on 2026-10-03: a grep of the entire init surface (`/etc/cont-init.d`, `/plex-common.sh`,
+`/usr/local/bin`) for `ManualPortMapping` and `PLEX_PREFERENCE` returned **NONE**. The image writes
+only six preferences, and this is not one of them. It is the same class of manual last mile as the
+`HardwareAcceleratedCodecs` toggle in `runbooks/plex-quicksync.md`.
+
+In the Plex web UI (`https://plex.home.keithjackson.dev`, on the LAN):
+
+**Settings → Remote Access**
+- Ensure **Enable Remote Access** is on. (`PublishServerOnPlexOnlineKey` was already `1`.)
+- Tick **Manually specify public port** and set it to **32400**.
+- **Apply**, then let the panel settle. It should report *"Fully accessible outside your network"*.
+
+Without this tick Plex waits for a NAT mapping it can never obtain and publishes no remote
+connection, so the port-forward in step 1 sits there unused.
+
+## How to verify
+
+Do these in order; each one isolates a different half.
+
+**1. The port is open from outside.** From a phone on cellular, or any off-LAN host:
+
+    curl -sk -o /dev/null -w '%{http_code}\n' https://136.51.115.252:32400/identity
+
+`200` means the forward and the policy are both good. A hang or refusal means step 1.
+
+**2. Plex itself believes it is reachable.** Settings → Remote Access must say *"Fully accessible
+outside your network"*. If it says *"not available outside your network"* while test 1 passes, step 2
+was not applied.
+
+**3. plex.tv is handing out a remote connection.** The decisive test, because it is what clients
+actually consume: sign in to [app.plex.tv](https://app.plex.tv) from an off-LAN network and confirm
+the server appears and plays. A client finds this server through plex.tv's connection list, never
+through DNS.
+
+**4. The policy is admitting the traffic.** If test 1 fails *after* the forward is confirmed, the
+CiliumNetworkPolicy is the next suspect:
+
+    ./infra/devcontainer/dev.sh run bash -c \
+      'kubectl --context k3s-prod -n plex get cnp plex-allow-remote-32400'
+
+See `runbooks/denied-flows.md` for reading actual drops.
+
+## Why no alert
+
+Every in-cluster signal is green whether or not remote access works. The Service has endpoints, the
+pod is `Running`, the policy is applied, and `plex-down` watches
+`kube_deployment_status_replicas_available` — none of which can see a router with a missing
+port-forward or an unticked checkbox. An alert would need an off-LAN prober, which this cluster does
+not have. **Test 3 above is the only honest check, and it is manual.**
+
+## Why not the Gateway URL (`ADVERTISE_IP`)
+
+The alternative was to leave Plex on `ClusterIP` and set
+`ADVERTISE_IP: https://plex.home.keithjackson.dev:443`, which writes Plex's `customConnections`
+preference and would have put the server on plex.tv's list via the existing L7 Gateway — no new WAN
+hole. That mechanism was verified to work on an already-claimed server (the `ADVERTISE_IP` block at
+line 126 of the image's `40-plex-first-run` sits **outside** the claim guard, and `/.firstRunComplete`
+lives on the overlay FS, not the `/config` PVC, so it re-runs every pod start), and the Gateway path
+itself was measured serving Plex correctly: `HTTP 200`, `claimed="1"`, with `401` on every protected
+endpoint.
+
+**The operator chose the native path anyway (2026-10-03), and the reasons are real:** Plex terminates
+its own TLS with its `*.plex.direct` certificate, and video streams bypass the Envoy L7 hop instead
+of being proxied through it. The cost is a new unfiltered WAN port — acceptable because Plex
+authenticates every caller itself. `ADVERTISE_IP` therefore stays unset in `catalog/plex.yaml`, and
+the deferral note there is still accurate.
+
+## Known wart: remote clients are classified as LAN
+
+`apps/wave-0/plex-remote.yaml` uses `externalTrafficPolicy: Cluster`, which SNATs the client to the
+ingress node's address (`192.168.2.1x`). That falls inside the `192.168.0.0/16` entry of
+`ALLOWED_NETWORKS`, so **Plex treats every remote stream as local.** Measured on the Gateway path,
+which SNATs the same way: Plex logged `[10.0.2.214:36040 (Allowed Network (WAN))]`.
+
+This is a **metering** effect, not a security one — authorization is unchanged, as the 401s above
+show — but Plex's remote-quality and bandwidth limits will not bind on these streams.
+
+`Local` would preserve the client IP and fix it. **Do not switch to it.** Cilium's L2 announcer
+elects one lease holder per Service and does *not* follow the backend pod, so the VIP can be
+announced by a node with no Plex pod, black-holing the port while everything reports healthy. That
+exact change cost this LAN an 8-day Minecraft outage in September 2026 — the full write-up is at the
+top of `platform/crafty/service-l4.yaml`. Plex is a single-replica `Recreate` Deployment, so the
+node mismatch is the normal case here, not a rare one.
@@ -15,7 +15,7 @@ docs/13 exactly. Do not diverge from it here — escalate spec problems, don't f
 | `servicedefinition.schema.json` | Formal JSON Schema (draft 2020-12) for a ServiceDefinition. Enforces enums, required fields, patterns, and cross-field invariants. |
 | `renderer-contract.md` | Authoritative field → rendered-resource mapping, with requirement IDs. The contract other workstreams consume. |

-**Schema version:** 1.0.0 (in the schema `description`). Treat any change as an interface change:
+**Schema version:** 1.8.0 (in the schema `description`, which carries the full per-version changelog — this line had said `1.0.0` since 1.1.0 and is now synced). Treat any change as an interface change:
 propose on the task board, update dependent renderers/tests in the same change, never silently break a
 consumer (coordination-protocol §2).

@@ -35,6 +35,9 @@ consumer (coordination-protocol §2).
 - `ingressRoutes[].host` must match `<label>.home.keithjackson.dev`.
 - `storage[].class ∈ {longhorn, bulk}`; `exposure ∈ {internal, public}`; `egress ∈ {internet, vpn}`;
  `auth.mode ∈ {oidc, forward-auth, native, none}`.
+- `hardware.gpu ∈ {intel-quicksync}` — a host **capability class**, never a device path. The enum
+  is the contract: the mechanism that satisfies it lives only in `catalog/render.py` (see
+  `renderer-contract.md` §`hardware`).
 - `additionalProperties: false` everywhere — unknown keys are rejected (catches typos in review/CI).

 ## Validate examples against the schema — L1, no cluster (R-TEST-02)
@@ -23,6 +23,7 @@ platform must render, and cites the requirement each mapping satisfies. Changing
 | `dependencies` | `[]` → reaches **only DNS** (default-deny) | R-NET-01 |
 | `inClusterClients` | `[]` → no direct east-west client; only the Gateway may reach an exposed port | R-NET-01/02 |
 | `auth.mode` | `forward-auth` for any exposed UI | R-IDP-02 |
+| `auth.apiRoutes` | `[]` → the sidecar renders **no** `--api-route`; every unauthenticated request is 302'd to Keycloak | R-IDP-02 |
 | `update.channel` | `stable` | R-UPD-01 |
 | `update.autoRollbackOnFailedHealth` | `true` | R-UPD-03 |
 | `observability.metrics` | `true` | R-OBS-01 |
@@ -160,6 +161,158 @@ platform must render, and cites the requirement each mapping satisfies. Changing
 - **Requirements:** **R-STATE-01** (rebuild survival), **R-STATE-02** (Longhorn replication),
  **R-STATE-04** (bulk tier). (docs/04)

+### `availability` (`replicas`) — workload kind, per-replica volumes, spreading, budget
+
+**Schema 1.10.0.** Availability INTENT. It names no Kubernetes object, no topology key and no
+budget — exactly like `hardware.gpu`, and for the same reason: the mechanism below should be
+replaceable without a catalog edit.
+
+- **Absent, or `replicas: 1`** → a **Deployment**, standalone PVC(s), `strategy: Recreate` when there
+  is storage, no PodDisruptionBudget. Byte-identical to pre-1.10.0 output, so every existing app is
+  untouched.
+- **`replicas: N > 1`** → all of the following, together:
+  | Rendered | Why |
+  |---|---|
+  | **StatefulSet** instead of Deployment | A longhorn PVC is **ReadWriteOnce**. One RWO claim cannot attach on two nodes, so a multi-replica Deployment does not degrade — the second pod never schedules. |
+  | longhorn storage becomes **`volumeClaimTemplates`** | Gives each replica its own claim (`<vol>-<svc>-<ordinal>`, e.g. `config-pihole-0`). The template's `metadata.name` **must** equal the pod volume name the container mounts. The standalone PVC is then **not** emitted — it would be an orphan nothing mounts. |
+  | `bulk` storage stays a **standalone PVC** | A bulk entry is a static, pre-bound NFS PV over data that already exists outside the cluster and is shared read-only. Per-replica copies are neither possible nor wanted. |
+  | **required** `podAntiAffinity` on `kubernetes.io/hostname` | `preferred` would let every replica land on one node and deliver none of the intended node-failure tolerance, while every object still reported healthy. |
+  | `podManagementPolicy: Parallel` | `OrderedReady` serialises first-boot work across replicas; `apps-wave-0` has `timeout: 5m`. There is no ordering requirement between replicas to protect. |
+  | **PodDisruptionBudget**, `minAvailable: 1` | Stops a *voluntary* disruption (node drain, upgrade) taking the last replica. It does nothing about a node dying — that is what the replica count and anti-affinity are for. Never rendered at `replicas: 1`, where `minAvailable: 1` would block every drain of that node forever. |
+  | `observability.livenessWorkload.kind` defaults to **`StatefulSet`** | The `up` SLI reads kube-state-metrics by workload kind. A Deployment-shaped rule against a StatefulSet `absent()`s forever — an alert that can never fire. Deriving the default means bumping the replica count cannot leave a stale rule behind. |
+
+**Two costs, both deliberate, neither hidden:**
+
+1. **N replicas need N schedulable nodes.** On a 3-node cluster, losing two nodes leaves a replica
+   `Pending`; since `apps-wave-0` is `wait: true` and `apps-wave-1` `dependsOn` it, that **stalls the
+   app tier**. That coincides with having lost etcd quorum on a 3-node control plane, so nothing is
+   reconciling at that point anyway — the stall is a symptom, not a new failure mode.
+2. **Only declare `> 1` for a service whose on-disk state is reproducible from git.** A replica with
+   a fresh volume comes up serving whatever it can bootstrap itself and nothing else. For pihole that
+   prerequisite was done first — the admin password and the adlists were declared in
+   `decisions/2026-10-04-pihole-declare-state.md` — because otherwise a second replica would have
+   served DNS while blocking nothing, with a different password.
+
+**Requirements:** R-PLAT-02 (the workload), R-STATE-01/02 (per-replica replicated volumes),
+R-OBS-01 (the SLI must follow the rendered kind).
+
+### `hardware` (`gpu`) — host device capability
+- **The field is INTENT, never MECHANISM.** `hardware.gpu` is an **enum of capability classes**
+  (today: `intel-quicksync`), not a device path, not a hostPath, not a resource request. The
+  ServiceDefinition says *what class of host device this workload needs*; how the platform hands it
+  over is `catalog/render.py`'s business and lives nowhere else. **Do not "simplify" this into a raw
+  `/dev/...` path.** The payoff is no longer hypothetical: on **2026-10-01** the mechanism was
+  swapped outright — from a `/dev/dri` hostPath to the Intel GPU device plugin's extended resource —
+  as a **`catalog/render.py` change plus a re-render, with ZERO catalog edits** and no re-review of
+  per-service device wiring. `catalog/plex.yaml`'s `hardware:` block came through byte-identical.
+  Extend the enum for a genuinely different capability class (e.g. `nvidia-nvenc`); never to name a
+  second route to the same one.
+- `gpu: intel-quicksync` → on the **app container only** (never a sidecar), **exactly one thing**:
+
+      resources:
+        limits:
+          gpu.intel.com/i915: 1
+
+  **That is the whole mapping.** No volume, no volumeMount, no `securityContext`, no `nodeSelector`,
+  no privileged anything.
+  - **`limits`, not `requests`, and that is not a style choice.** An extended resource must be
+    expressed in `limits`; Kubernetes mirrors `requests` from `limits` automatically, and a
+    requests-only extended resource is rejected as invalid.
+  - The resource is advertised by the **`intel-gpu-plugin` DaemonSet** (`platform/intel-gpu-plugin/`,
+    its own Flux Kustomization at `clusters/base/intel-gpu-plugin.yaml`, pinned
+    `intel/intel-gpu-plugin:0.37.0`). **Measured 2026-10-01:** all three prod nodes report
+    `.status.allocatable["gpu.intel.com/i915"] == 1`; DaemonSet 3 desired / 3 ready / 3 available.
+  - This is the **only** app-container `resources` render path in the file, and it is deliberately
+    narrow. **Do not generalise it into a CPU/memory resources feature** while passing through — that
+    is a separate field with a separate justification, and `render.py --selftest` asserts the emitted
+    dict *exactly* so the attempt fails loudly.
+- **WHY THE hostPath MECHANISM WAS ABANDONED — read this before "simplifying" it back.** The first
+  implementation (merged 2026-09-30) rendered a `dri` hostPath volume at `/dev/dri` plus a read-write
+  mount. It was **measured not to work** in the live Plex pod on k3s-prod-2 on **2026-10-01**. The
+  device node was **visible** in the container — `crw-rw---- root video2 226,128 renderD128` — and
+  **could not be opened**:
+
+  | attempt | result |
+  |---|---|
+  | `uid 0`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+  | `uid 1000`, `O_RDWR` | **`EPERM`** "Operation not permitted" |
+
+  `EPERM` is the **device cgroup** refusing, not `EACCES` from file permissions — the two look
+  identical from the outside and are not the same problem. Proven three ways: root on the **host**
+  opens the same device `rc=0`; the node is unified `cgroup2fs`; and the container's runc
+  `config.json` reads `linux.resources.devices: [{allow: False, access: 'rwm'}]` with
+  `linux.devices: None` — **one deny-all rule, no allow entry for major 226**. A hostPath is a
+  *filesystem* operation: it binds the device node in but adds **no cgroup allow rule**, because
+  Kubernetes cannot infer one from a volume. **Note that `uid 0` failed too** — root does not bypass
+  the cgroup device gate, so "run it as root" is not a workaround. Neither is `privileged: true`:
+  Kyverno `restrict-host-namespaces` is **Enforce** on app namespaces and rejects it.
+  The device plugin is what produces the allow rule — confirmed in the v0.37.0 source,
+  `createDeviceSpecsFromDrmFiles` emits `DeviceSpec{…, Permissions: "rw"}`, which the kubelet turns
+  into both the device injection and the matching cgroup allow.
+- **THE FAILURE MODE THIS TRADED INTO IS WORSE IN BLAST RADIUS, AND IT IS UNAVOIDABLE.** Stated
+  plainly rather than sold:
+
+  | | device/resource unavailable ⇒ |
+  |---|---|
+  | **Before** (hostPath, `DirectoryOrCreate`) | kubelet mounts an empty dir; Plex runs and **silently software-transcodes**. Invisible, but harmless to everything else. |
+  | **Now** (extended resource) | the pod is **Unschedulable**. `apps-wave-0` is `wait: true` and `apps-wave-1` `dependsOn` it, so this **stalls the app tier**. |
+
+  There is **no graceful-degradation variant**: an extended resource is a hard scheduling
+  constraint, and it is also the only thing that produces the cgroup allow rule. The workload's
+  schedulability is now **coupled to the `intel-gpu-plugin` DaemonSet** — a plugin that is broken,
+  removed, or merely not yet reconciled takes the app tier with it. That is the price of hardware
+  transcoding working at all.
+  - **Escape hatch, one field and no code change:** delete the `hardware:` block from the
+    ServiceDefinition (`catalog/plex.yaml`) and re-render. The workload returns to software
+    transcoding and the tier unblocks. `runbooks/plex-quicksync.md` carries it as a first-class
+    procedure, because it is what someone will need at 2am.
+  - The old mechanism's cost is **gone**, and that is the one improvement: a missing device is no
+    longer silent. There is still no metric and no alert (media apps here render no
+    `metricsEndpoint`, so no ServiceMonitor exists), but "Plex is Unschedulable" is loud, whereas
+    "Plex is quietly software-transcoding" never was.
+- **`podSecurityContext.supplementalGroups` is deliberately NOT rendered, and adding it would be a
+  regression rather than a hardening.** This survives the mechanism swap unchanged, and it was
+  re-measured against the plugin's behaviour. `renderD128` is `root:render` `0660`, so the obvious
+  move is to push the host `render` GID in — and it does not work. **Measured in the running
+  container on prod-2 (2026-09-30, re-measured 2026-10-01):** the image's own
+  `/etc/cont-init.d/45-plex-hw-transcode-and-connected-tuner` hook `find`s every char device under
+  `/dev/dri`, `stat -c '%g'` each one's GID, `groupadd`s a matching group *inside the container* when
+  none exists, and `usermod -a -G` the app user into it — every start, nothing hardcoded. It does
+  exactly the same thing to a **plugin-injected** device as to a bind-mounted one, which is why this
+  swap needed no group work. `cont-init.d` completes before `services.d`, so the account database is
+  already correct when the service execs. The service then starts via `exec s6-setuidgid plex …`, and
+  `s6-setuidgid` **replaces** the process group list from `/etc/group` rather than appending to it —
+  measured live, `Plex Media Server` ran with `Groups: 44 100 992 1000`, where **992 is the host
+  `render` GID**, while root's gid `0` (which PID 1, `s6-svscan`, does hold) was **gone**. A
+  Kubernetes `supplementalGroups` therefore reaches `s6-svscan` and is discarded before the app ever
+  starts. It would be inert *and* it would mislead whoever debugs this next. The renderer asserts its
+  own absence in `render.py --selftest`.
+  - A corollary worth having: because the hook re-`stat`s the device on every start, the host GID
+    **drifting** (a node rebuild renumbering the *systemd dynamic* `render` group) is self-healing.
+    No GID is pinned anywhere in this repo.
+  - **Scope of that measurement, stated so nobody over-reads it.** "The group is the image's job" was
+    established for an image that resolves the device group itself. A FUTURE consumer that runs
+    non-root and has no equivalent init hook would still need the group solved somehow — and the
+    answer for that case is a **renderer** change (per-image handling), not a new ServiceDefinition
+    field. The enum is what keeps that option open, and the 2026-10-01 swap is the proof it works.
+    The first such consumer must re-measure rather than assume this bullet covers it.
+- **Deliberately NOT rendered** from this field: no `privileged: true` (Kyverno
+  `restrict-host-namespaces` is Enforce on app namespaces and would reject it, and it would not have
+  helped anyway — see the `EPERM` table); no `nodeSelector` (placement is a separate concern with its
+  own justification — this field is about the device, not about where the pod lands; the extended
+  resource does constrain placement to nodes the plugin advertises, but that is the **scheduler's**
+  doing, not a selector this renderer writes); no CPU/memory requests or limits (see above — the
+  `resources` block carries the extended resource and nothing else).
+- **Historical, kept so a reader recognises it rather than reintroduces it:**
+  `type: DirectoryOrCreate` on the retired hostPath was load-bearing *for that mechanism* — it is
+  what made a GPU-less node degrade to software transcoding instead of failing the mount and
+  stalling the app tier. That rationale died with the hostPath; it is **not** an argument for bringing it back, since
+  the mechanism it protected never granted access in the first place.
+- Absent ⇒ **nothing at all** is rendered; every service that does not opt in stays byte-identical
+  (no `resources` key, no volume, no mount, no `securityContext`).
+  Ignored for `kind: vm` — a VM receives devices through KubeVirt's own `domain.devices`.
+- **Requirements:** **R-PLAT-02** (the PaaS provides the host capability; the app declares need).
+
 ### `ingressRoutes[]` (`host` / `port`) — HTTPRoute + Certificate + DNS
 - Each entry renders an **HTTPRoute** attaching `host` → the service `port` on the shared Gateway.
 - TLS: the Gateway terminates with the cert-manager **wildcard `*.home.keithjackson.dev`** certificate
@@ -202,7 +355,7 @@ platform must render, and cites the requirement each mapping satisfies. Changing
  undeclared), **R-GITOPS-02** (the allowance is authored in the one reviewable object rather than
  surviving as a live-applied cluster object). Auditable via Hubble (R-SEC-03, R-OBS-04).

-### `auth` (`mode` / `groups[]`) — identity integration (Keycloak / oauth2-proxy)
+### `auth` (`mode` / `groups[]` / `apiRoutes[]`) — identity integration (Keycloak / oauth2-proxy)
 - `oidc` → register the app as a **Keycloak OIDC client** (declarative realm/client config, GitOps;
  client secret SOPS-encrypted). App does its own login redirect.
 - `forward-auth` → protect the service's `ingressRoutes` at the **Gateway** with **oauth2-proxy →
@@ -213,6 +366,20 @@ platform must render, and cites the requirement each mapping satisfies. Changing
 - `groups[]` → the Keycloak groups permitted access, enforced by the OIDC client mappers /
  oauth2-proxy `allowed_groups`. **Schema requires ≥1 group for `oidc`/`forward-auth`** — access is
  "in an authorized group," not merely "logged in."
+- `apiRoutes[]` (**`forward-auth` only**) → one repeatable **`--api-route=<regex>`** per entry on
+  the per-app oauth2-proxy sidecar, making matching paths answer **401** instead of a **302** to
+  Keycloak. Default `[]` renders **no flag at all**, which is the common case. It exists because
+  oauth2-proxy's built-in AJAX guard fires only on an exact `Accept: application/json`, which
+  these apps' JavaScript does not consistently send: the XHR is 302'd, the browser follows it
+  invisibly, the live SSO session satisfies it silently, and the call resolves 200 carrying the
+  app's **HTML** where the JavaScript expected JSON — and a POST following a 302 is re-issued as
+  a GET with its **body dropped**. It is **per-service and must never become a renderer
+  constant**: the path shape belongs to the app (radarr `/signalr/`, jackett `/api/`, pihole
+  none), and a wrong constant does not merely fail to help — a path a **browser navigates to**,
+  answered 401, shows a logged-out operator a bare error page and never offers a sign-in. Each
+  entry must be **anchored at `^/`** and continue with literal path characters; `^/`, `^/.*`,
+  `.*`, an unanchored `/api/` and any alternation are **rejected** by the schema and by the
+  renderer (alternation is refused rather than supported because the flag is repeatable).
 - Keycloak, oauth2-proxy, and app→Keycloak / proxy→Keycloak / Keycloak→DB flows are ordinary
  workloads under default-deny with explicit allows (ties back to `dependencies`).
 - **Requirements:** **R-IDP-02** (protect OIDC-capable and bare UIs), **R-IDP-06** (group-based
@@ -271,6 +438,12 @@ platform must render, and cites the requirement each mapping satisfies. Changing
 ## Cross-field invariants the renderer / L0-L1 tests rely on
 1. **kind ↔ source:** container⇒`image`, vm⇒`vmTemplate` (schema-enforced).
 2. **auth ↔ groups:** `oidc`/`forward-auth` ⇒ non-empty `groups` (schema-enforced; R-IDP-06).
+2a. **auth ↔ apiRoutes:** a non-empty `auth.apiRoutes` ⇒ `auth.mode: forward-auth`
+   (schema-enforced). It renders onto the per-app oauth2-proxy sidecar, and no sidecar exists for
+   `oidc`/`native`/`none`, so declaring it elsewhere would be a **silent no-op** — the operator
+   believes the app's XHRs now answer 401 and nothing changed. Every entry must be anchored at
+   `^/` (schema-enforced **and** enforced in `catalog/render.py`, so a breadth mistake is caught by
+   `render-golden` even on a run where the schema stage did not execute).
 3. **default-deny is unconditional:** every namespace gets the deny baseline regardless of
   `dependencies`; `dependencies` only *adds* allows (R-NET-01/02). The same holds for
   `inClusterClients`: it adds one namespace+port ingress allow and never relaxes the baseline.
@@ -286,6 +459,18 @@ platform must render, and cites the requirement each mapping satisfies. Changing
   `external-dns.alpha.kubernetes.io/controller: none`, and no `public` service's route carries it.
   Derived from `exposure` alone; enforced on every gate run by
   `conformance/tests/l0-static/internal-services-publish-no-public-dns` (R-DNS-03).
+8. **`hardware` is inert until opted into, and its mechanism is renderer-local — now DEMONSTRATED,
+   not merely claimed.** An SD with no `hardware` key renders no `resources` block, no volume and no
+   volumeMount; an SD *with* it gains a `gpu.intel.com/i915: 1` **limit on the app container only**
+   (never a sidecar) and no `securityContext` either. All directions are proven by
+   `render.py --selftest`, and each assertion is mutation-tested. The resource name
+   `gpu.intel.com/i915` appears in `catalog/render.py` and nowhere else in the catalog path, so the
+   capability's implementation can be replaced without touching a single ServiceDefinition — which
+   is exactly what happened on **2026-10-01**, when the mechanism moved from a `/dev/dri` hostPath
+   (measured non-functional: device-cgroup `EPERM`) to the device plugin's extended resource with
+   **zero** catalog edits. The cost of that mechanism is recorded in §`hardware`: the workload is now
+   **Unschedulable** when the resource is absent, which stalls the app tier, and the escape hatch is
+   to delete the `hardware:` block and re-render.

 ## Requirement traceability (contract-level)
 Primary **R-GITOPS-02** (single reviewable object wires every requirement). Field-level: R-NET-01/02
@@ -2,7 +2,7 @@
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://home.keithjackson.dev/schemas/servicedefinition.schema.json",
  "title": "ServiceDefinition",
-  "description": "Operator-facing onboarding contract for the home k3s + Longhorn platform. One ServiceDefinition per file under catalog/. Captures INTENT; the platform compiles it into concrete resources (see schemas/renderer-contract.md). Canonical source of truth: spec docs/13-service-catalog.md. Serves R-GITOPS-02; ties into R-NET-*, R-STATE-*, R-TLS-*, R-DNS-*, R-OBS-*, R-UPD-*, R-IDP-*. Schema version: 1.7.0 (no field changes; a MAPPING change to `exposure`, which had never had an effect on DNS. `exposure: internal` now renders ExternalDNS's own controller annotation onto every HTTPRoute it emits, set to a value other than `dns-controller`, so external-dns (v0.21.0, source/gateway.go -> annotations.IsControllerMismatch) skips the route before resolving any hostname and the service is INCAPABLE of producing a public record. Both docs/09 R-DNS-03 and schemas/renderer-contract.md already said `internal` renders no public DNS record; nothing enforced it, and six internal services -- jackett, pihole, qbittorrent, radarr, semaphore, sonarr -- were publishing `<name>.home.keithjackson.dev` into the PUBLIC Cloudflare zone answering with the RFC1918 Gateway VIP, which published the internal subnet and a full service inventory and made each name a DNS-rebind primitive. Enforced on every gate run by conformance/tests/l0-static/internal-services-publish-no-public-dns). 1.6.0 (splits `observability.metrics` from WHERE metrics live: adds optional `metricsEndpoint` (port/name/path/interval) and optional `livenessWorkload` (kind/name). `metrics: true` previously rendered a ServiceMonitor hardcoded to a port named `metrics` that no rendered Service ever defined, so all twelve app monitors discovered ZERO targets and every `up{job=...}` SLI was structurally inert on prod; a monitor is now rendered only from a declared `metricsEndpoint`, which also renders the Service port and the monitoring-namespace ingress allow, and the `up` SLI is expressed against kube-state-metrics -- a signal that exists for every workload without a per-app exporter. `livenessWorkload` names the object whose readiness IS the service, for the case where the running workload is not the one this SD renders). 1.5.0 (adds optional `inClusterClients` — additional in-cluster clients, named by NAMESPACE, admitted to a declared port of this service DIRECTLY, bypassing the shared Gateway. `ingressRoutes` admits only the Gateway proxy identity and `dependencies` only links catalog services, so a PLATFORM controller dialling a Service ClusterIP had no authorable form at all; that gap is what deadlocked prod's GitOps when its GitRepository moved to the in-cluster Gitea Service. Renders one least-privilege ingress CiliumNetworkPolicy per entry — the named namespace and the declared port, nothing wider). 1.4.0 (makes the `bulk` tier REAL: a bulk storage entry now carries `export` (the absolute NFS export path on the ${BULK_NFS_SERVER} host) and renders a STATIC, pre-bound NFS PersistentVolume + PVC over data that already exists outside the cluster, plus optional `subPath` and `readOnly` on any entry. Resolves the OQ-2 bulk-backing stub for STATIC pre-existing datasets; dynamic bulk provisioning remains unimplemented and unneeded). 1.3.0 added optional pod-level podSecurityContext for the container kind — k8s PodSecurityContext emitted verbatim onto the pod spec; its fsGroup unblocks non-root images e.g. semaphore/UID-1001 writing a mounted Longhorn PVC). 1.2.0 added optional per-container env/envFrom/lifecycle for the container kind — k8s EnvVar/EnvFromSource/Lifecycle shapes emitted verbatim onto the app container; unblocks headless-install/DB-backed apps e.g. gitea/nextcloud/semaphore/unifi/minecraft. 1.1.0 added the shape invariant exposure:public => auth.mode!=none (docs/13 v1.1 / docs/15 / R-IDP-02).",
+  "description": "Operator-facing onboarding contract for the home k3s + Longhorn platform. One ServiceDefinition per file under catalog/. Captures INTENT; the platform compiles it into concrete resources (see schemas/renderer-contract.md). Canonical source of truth: spec docs/13-service-catalog.md. Serves R-GITOPS-02; ties into R-NET-*, R-STATE-*, R-TLS-*, R-DNS-*, R-OBS-*, R-UPD-*, R-IDP-*. Schema version: 1.10.0 (adds optional `availability` for the container kind, one member: `availability.replicas`. It is AVAILABILITY INTENT, not a scheduling primitive -- it names no Kubernetes object, no topology key and no budget. 1 (the default) renders a Deployment exactly as before, so every pre-existing app stays byte-identical; >1 means the service must survive the loss of a single node, and the renderer satisfies that with a StatefulSet whose longhorn volumes become per-replica volumeClaimTemplates (a longhorn PVC is ReadWriteOnce and cannot attach on two nodes, so a multi-replica Deployment is UNSCHEDULABLE, not merely suboptimal), REQUIRED podAntiAffinity on kubernetes.io/hostname (preferred would let every replica land on one node and deliver none of the intended tolerance while reporting healthy), podManagementPolicy Parallel, a PodDisruptionBudget of minAvailable 1, and an `up` SLI that follows the rendered kind (a Deployment-shaped rule against a StatefulSet absent()s forever and can never fire). Added for pihole, where LAN DNS -- a critical network service -- was one pod on one node: measured 2026-10-04, losing the node running it meant roughly 6-10 minutes with no LAN resolution (tolerationSeconds 300 plus node-NotReady detection plus a Longhorn RWO force-detach). Deliberately NOT usable until a service's state is reproducible from git, because a replica with a fresh volume otherwise comes up serving but unconfigured -- for pihole that prerequisite was done first (decisions/2026-10-04-pihole-declare-state.md). Capped at 5: a count above the node count can never schedule under the anti-affinity this implies. Mechanism documented in schemas/renderer-contract.md `### availability`). 1.9.0 (adds optional `auth.apiRoutes` -- a per-service list of anchored path regexes the forward-auth oauth2-proxy sidecar answers with 401 instead of a 302 to Keycloak, rendered as one repeatable `--api-route` per entry. The `auth` block stays `additionalProperties: false`; this is one named, documented key, not an opened door. Without it there was NO per-app override hook at all, so the only way to answer an app's XHRs with 401 was a renderer-wide constant -- and the path shape is a property of the app: jackett's is `/api/` (33 x 302 vs 2 x 401 measured 2026-09-30, 4 of 6 POSTs answered by the proxy and never reaching the app), radarr's is `/signalr/` (232 x 302 measured 2026-10-01, every one a POST from a left-open browser tab that is never re-prompted), and pihole serves no `/api/` at all. Each entry must be anchored at `^/` and continue with literal path characters, because a too-broad regex turns a plain page load into a 401 and the service looks permanently broken to a logged-out user. Enforced on every gate run by ci/l0.sh's schema stage over catalog/*.yaml, by catalog/render.py refusing to render an unanchored entry, and by conformance/tests/l0-static/oauth2-proxy-api-routes-per-app)). 1.8.0 (adds optional `hardware` - a host HARDWARE CAPABILITY declaration for the container kind, currently one member: `hardware.gpu: intel-quicksync`, meaning the workload needs the host iGPU render device for hardware video transcode. It is an ENUM OF INTENT, not a device path: the renderer satisfies it today with just two things -- a /dev/dri hostPath volume (`DirectoryOrCreate`) and a read-write /dev/dri volumeMount on the app container -- and a later move to the Intel GPU device plugin is then a render.py change with no catalog edit. It sets NO podSecurityContext.supplementalGroups: measured in-container on prod 2026-09-30, the pms-docker cont-init hook adds the app user to the device's group inside the container and `s6-setuidgid` then REPLACES the process group list from /etc/group, discarding anything Kubernetes supplied, so such an entry would be inert and misleading. Renders NOTHING when absent, so every existing app stays byte-identical. Emits no privileged securityContext (Kyverno restrict-host-namespaces would reject it and it is not needed), no nodeSelector (node placement is a separate concern with its own justification) and no resource requests). 1.7.0 (no field changes; a MAPPING change to `exposure`, which had never had an effect on DNS. `exposure: internal` now renders ExternalDNS's own controller annotation onto every HTTPRoute it emits, set to a value other than `dns-controller`, so external-dns (v0.21.0, source/gateway.go -> annotations.IsControllerMismatch) skips the route before resolving any hostname and the service is INCAPABLE of producing a public record. Both docs/09 R-DNS-03 and schemas/renderer-contract.mdLine truncated
  "type": "object",
  "additionalProperties": false,
  "required": ["name", "source"],
@@ -139,6 +139,34 @@
        "supplementalGroups": { "type": "array", "items": { "type": "integer" }, "description": "Additional GIDs added to the first process in each container." }
      }
    },
+    "availability": {
+      "type": "object",
+      "description": "Optional availability intent for the container kind: how many replicas of this workload must be running. INTENT ONLY -- it names no Kubernetes object and no scheduling primitive. How the platform satisfies it (workload kind, per-replica volumes, node spreading, disruption budget) is catalog/render.py's business, documented in schemas/renderer-contract.md `### availability`. Absent => replicas 1 and a Deployment, so every pre-existing app renders byte-identical. Ignored for kind=vm (a VM's availability is a different problem with different primitives).",
+      "additionalProperties": false,
+      "minProperties": 1,
+      "properties": {
+        "replicas": {
+          "type": "integer",
+          "description": "How many replicas must be running. 1 (default) renders a Deployment exactly as before. >1 declares that the service must survive the loss of a single node; the renderer then emits a StatefulSet with per-replica volumeClaimTemplates, required per-node anti-affinity, podManagementPolicy Parallel and a PodDisruptionBudget. Only declare >1 for a service whose on-disk state is reproducible from git -- a replica with a fresh volume comes up serving whatever it can bootstrap itself, and nothing else.",
+          "minimum": 1,
+          "maximum": 5,
+          "default": 1
+        }
+      }
+    },
+    "hardware": {
+      "type": "object",
+      "description": "Optional host HARDWARE CAPABILITY the workload needs, for the container kind (docs/13). This field declares INTENT (`what device class this service needs`), never MECHANISM (`which host path to bind-mount`, `which extended resource to request`). That separation is the whole point of the enum: the platform currently satisfies `intel-quicksync` with a /dev/dri hostPath volume and a matching mount on the app container, but a later migration to the Intel GPU device plugin (`gpu.intel.com/i915` resource requests, no hostPath) must be a catalog/render.py change and NOTHING else - no ServiceDefinition may need editing, and no reviewer may have to re-derive per-service device wiring. Do NOT 'simplify' this into a raw device path, a hostPath, or a resource request; that would push a platform implementation detail into the operator-facing contract and make the migration a catalog-wide edit. Absent => no device wiring at all (existing apps render byte-identical). Ignored for kind=vm (a VM gets devices through KubeVirt's own domain.devices).",
+      "additionalProperties": false,
+      "minProperties": 1,
+      "properties": {
+        "gpu": {
+          "type": "string",
+          "description": "GPU capability class this service needs. `intel-quicksync` = the host's Intel Quick Sync Video render device (the iGPU's /dev/dri/renderD* node) for hardware video transcode/decode. Enum, not a free-form device path, so the satisfying mechanism stays swappable (see the parent description). Extend the enum when a genuinely different capability class appears (e.g. nvidia-nvenc), never to name a different path to the same one.",
+          "enum": ["intel-quicksync"]
+        }
+      }
+    },
    "exposure": {
      "type": "string",
      "description": "internal = LAN-only (internal DNS + Gateway on LAN); public = additionally edge port-forwarded 80/443 (docs/09; R-DNS/R-NET).",
@@ -311,6 +339,16 @@
          "items": { "type": "string", "minLength": 1 },
          "uniqueItems": true,
          "default": []
+        },
+        "apiRoutes": {
+          "type": "array",
+          "description": "forward-auth ONLY: path regexes this app serves as a MACHINE API, which the oauth2-proxy sidecar must answer with 401 instead of a 302 to Keycloak (one --api-route per entry; the flag is repeatable and singular -- --api-routes does not exist in v7.8.1 and the proxy refuses to start with it). WHY THIS IS PER-APP AND NOT A RENDERER CONSTANT: oauth2-proxy's built-in AJAX guard fires only on an EXACT `Accept: application/json`, which these apps' JavaScript does not consistently send, so an unauthenticated XHR gets a 302 that the browser follows invisibly, the live Keycloak SSO session satisfies silently, and the call resolves 200 carrying the app's HTML where the JavaScript expected JSON -- no prompt, no error, and on a write it is worse, because a browser following a 302 re-issues a POST as a GET and DROPS THE BODY. Measured on jackett 2026-09-30: 33 x 302 against 2 x 401 on /api/ paths, with 4 of 6 POSTs answered by the proxy and never reaching the app (access-log upstream field `-`). Measured on radarr 2026-10-01, 23h after the cookie fix deployed: 232 x 302, ALL of them POST /signalr/messages/negotiate from the operator's own browser -- a left-open tab whose cookie the rename orphaned, polling forever, silently redirected every time and NEVER re-prompting. But the path shape is a property of the APP, not of the pattern: radarr's broken path is /signalr/ and jackett's is /api/, and pihole serves no /api/ at all, so a renderer-wide constant would apply one app's path shape to five. Declare only paths this app actually serves as JSON/RPC and that a browser never NAVIGATES to. BREADTH IS THE DANGER: too broad and a plain page load answers 401 instead of signing the operator in, i.e. the service looks permanently broken to a logged-out user and no sign-in is ever offered. The item pattern therefore REQUIRES each entry to be anchored at `^/` and to continue with literal path characters, which rejects `^/`, `^/.*`, `.*` and any unanchored or alternated regex; alternation is refused rather than supported because the flag is repeatable, so one regex per route costs nothing and keeps each entry readable. N/A for oidc/native/none (no sidecar is rendered).",
+          "items": {
+            "type": "string",
+            "pattern": "^\\^/[A-Za-z0-9._~/-]+(\\.\\*|\\$)?$"
+          },
+          "uniqueItems": true,
+          "default": []
        }
      }
    },
@@ -458,6 +496,26 @@
        }
      }
    },
+    {
+      "$comment": "auth.apiRoutes only has a consumer under forward-auth: it renders --api-route onto the per-app oauth2-proxy sidecar, and no sidecar is rendered for oidc/native/none. Declaring it elsewhere would be a SILENT NO-OP -- the operator believes the app's XHRs now answer 401 and nothing changed -- so it is rejected at the schema instead (R-IDP-02).",
+      "if": {
+        "required": ["auth"],
+        "properties": {
+          "auth": {
+            "required": ["apiRoutes"],
+            "properties": { "apiRoutes": { "minItems": 1 } }
+          }
+        }
+      },
+      "then": {
+        "properties": {
+          "auth": {
+            "required": ["mode"],
+            "properties": { "mode": { "const": "forward-auth" } }
+          }
+        }
+      }
+    },
    {
      "$comment": "public exposure must not be unauthenticated: an explicit auth.mode 'none' is disallowed on public services (docs/13 v1.1, docs/15, R-IDP-02). Absent auth defaults to forward-auth at render time, so it is permitted here.",
      "if": {