fix(cdi): single-replica scratch class + 40Gi UniFi disk (fix insufficient-storage import)

The UniFi Block import got past the device-permission fix but then failed
ReplicaSchedulingFailure / "insufficient storage": CDI sizes scratch ~= the disk
virtual size and, on the default 3-replica `longhorn` class, the ~60Gi scratch x3
(189G) on top of the disk's x3 (180G) overflowed the Longhorn tier (~125G/node).

- platform/longhorn/storageclasses.yaml: NEW `longhorn-scratch` StorageClass
  (numberOfReplicas: 1) — CDI scratch is ephemeral/throwaway, replicating it 3x is
  waste; if a node dies mid-import CDI just retries.
- platform/kubevirt/cdi-cr.yaml: spec.config.scratchSpaceStorageClass ->
  longhorn-scratch.
- apps/vms/unifi/20-datavolume.yaml: boot disk 60Gi -> 40Gi (UOS min is 25Gi;
  40Gi is comfortable for a home site and lighter on the replicated tier).

Co-Authored-By: Claude Opus 4.8 <[email protected]>

apps/vms/unifi/20-datavolume.yaml

@@ -12,8 +12,8 @@
 # STORAGE = longhorn-vm (the RWX/migratable StorageClass the storage workstream is adding):
 #   - accessModes: ReadWriteMany + volumeMode: Block  => REQUIRED for KubeVirt LIVE MIGRATION
 #     (evictionStrategy: LiveMigrate in the VM). A ReadWriteOnce disk cannot live-migrate.
-#   - 60Gi: the Ubuntu cloud image virtual size is ~3.5Gi; cloud-init growpart/growroot
+#   - 40Gi: the Ubuntu cloud image virtual size is ~3.5Gi; cloud-init growpart/growroot
-#     expands the root fs to fill the disk on first boot. 60Gi leaves ample room for the
+#     expands the root fs to fill the disk on first boot. 40Gi (UOS min is 25Gi) leaves room for the
 #     UniFi datastore + OS + podman images + updates.
 #
 # GOTCHA / FLAG: a Block RWX DataVolume import needs CDI scratch space and the importer pod
@@ -46,4 +46,4 @@ spec:
     storageClassName: longhorn-vm
     resources:
       requests:
-        storage: 60Gi
+        storage: 40Gi

platform/kubevirt/cdi-cr.yaml

@@ -18,9 +18,10 @@ spec:
   workload: {}
   config:
     # CDI's transient scratch space (used while converting/decompressing an imported image) is carved
-    # from the replicated Longhorn tier — the same StorageClass the final VM DataVolumes bind to, so
+    # from a dedicated SINGLE-replica Longhorn class — scratch is ephemeral/throwaway, so replicating
-    # there is no second storage tier to reason about. (docs/04, ADR-0003.)
+    # it 3x is waste and, for a large VM disk, the 3x scratch on top of the 3x final disk exhausts the
-    scratchSpaceStorageClass: longhorn
+    # Longhorn tier (ReplicaSchedulingFailure). See platform/longhorn/storageclasses.yaml. (docs/04, ADR-0003.)
+    scratchSpaceStorageClass: longhorn-scratch
     # Reserve headroom on the filesystem-mode PVC for qcow2 metadata so an import never wedges a volume
     # at 100%. 0.055 is CDI's documented default; pinned here so it is explicit and reviewable.
     filesystemOverhead:

platform/longhorn/storageclasses.yaml

@@ -75,6 +75,27 @@ parameters:
   # No fsType: migratable RWX volumes are Block-mode (raw device); a filesystem is never laid down
   # by Longhorn here — the guest owns the block device.
 ---
+# CDI scratch tier — SINGLE-replica, for CDI's EPHEMERAL import scratch space ONLY (ADR-0003).
+#
+# WHY a dedicated class: CDI carves transient scratch space (~= the imported disk's virtual size)
+# while converting/decompressing an image, then throws it away when the import finishes. Replicating
+# that 3x is pure waste — and for a large VM disk the 3x scratch (ON TOP of the 3x final disk) can
+# exhaust the Longhorn tier and fail the import with ReplicaSchedulingFailure / "insufficient
+# storage" (observed with the UniFi 60Gi import). One replica is correct for throwaway data: if the
+# node dies mid-import CDI just retries. Referenced by the CDI CR (spec.config.scratchSpaceStorageClass).
+apiVersion: storage.k8s.io/v1
+kind: StorageClass
+metadata:
+  name: longhorn-scratch
+provisioner: driver.longhorn.io
+allowVolumeExpansion: true
+reclaimPolicy: Delete            # scratch is ephemeral — always reclaim on PVC removal
+volumeBindingMode: Immediate
+parameters:
+  numberOfReplicas: "1"          # ephemeral throwaway data — no HA replication
+  staleReplicaTimeout: "30"
+  dataLocality: "best-effort"    # keep the lone replica local to the importer pod when possible
+---
 # Bulk media tier — large, NON-replicated data (Plex library, download staging, arr handoff).
 # docs/04 / R-STATE-04: 3x replicating multi-TB media across 3 mini-PCs is impractical, so bulk
 # media lives OUTSIDE the cluster and is consumed as an external NFS/SMB mount via this thin