fix(ci): de-flake the e2e job (loop-device image, fail-fast wait, diagnostics, hook timeout) (#621)

charts/platform/rook-ceph-cluster/templates/loop-device.yaml

@@ -4,6 +4,13 @@
 # DaemonSet creates a sparse image on the node and losetup-attaches it to a
 # FIXED loop device (storage.device / loopDevice.device) that the CephCluster
 # names explicitly. The container stays alive so the attachment persists.
+#
+# It runs on cephImage rather than something small like alpine because every
+# OSD in the cluster is blocked on this attachment: a registry the node has to
+# reach for this pod alone is a single point of failure, and a hung pull of one
+# surfaced half an hour later as an unrelated yucca-michael config error. This
+# image the node must pull anyway, and Tilt prepulls it the moment the cluster
+# exists. It carries losetup/truncate/mknod, so nothing has to be installed.
 apiVersion: apps/v1
 kind: DaemonSet
 metadata:
@@ -22,14 +29,13 @@ spec:
       hostPID: true
       containers:
         - name: loop-device
-          image: {{ .Values.loopDevice.image }}
+          image: {{ .Values.cephImage }}
           securityContext:
             privileged: true
           command: ["/bin/sh", "-c"]
           args:
             - |
               set -eu
-              apk add --no-cache util-linux >/dev/null 2>&1 || true
               IMG="{{ .Values.loopDevice.path }}"
               DEV="{{ .Values.loopDevice.device }}"
               mkdir -p "$(dirname "$IMG")"

charts/platform/rook-ceph-cluster/values.yaml

@@ -51,7 +51,6 @@ objectStore:
 # it; Rook's OSD then consumes it. Disable if your nodes have real disks.
 loopDevice:
   enabled: true
-  image: alpine:3.21@sha256:48b0309ca019d89d40f670aa1bc06e426dc0931948452e8491e3d65087abc07d
   sizeGiB: 20
   # where the backing image file lives on the node (under dataDirHostPath)
   path: /var/lib/rook/osd-loop.img