k8s-version-upgrade: unblock 1.34.9 — skip kubeadm CoreDNS addon + busybox-date fix

The 1.34.9 master upgrade hard-failed `kubeadm upgrade apply` preflight: CoreDNS
is at v1.12.4 (Keel auto-bumped it 1.12.1 -> 1.12.4 on 2026-05-26 via a stale
kube-system out-of-band annotation), and 1.12.4 is ahead of kubeadm 1.34.9's
bundled corefile-migration table ("start version not supported").

- scripts/update_k8s.sh: master `kubeadm upgrade apply` now runs with
  `--ignore-preflight-errors=CoreDNSMigration,CoreDNSUnsupportedPlugins
  --skip-phases=addon/coredns`. A dry-run proved --ignore ALONE would overwrite
  our custom split-horizon Corefile with kubeadm's default AND downgrade the
  image; --skip-phases leaves CoreDNS 100% untouched while the control plane
  upgrades. CoreDNS is pinned off Keel (keel.sh/policy=never) to stop the drift.
- stacks/k8s-version-upgrade/scripts/upgrade-step.sh: fix the preflight
  quiet-baseline (settle-window) check, which silently no-op'd on the ghcr
  claude-agent-service image's busybox `date` (can't parse ISO8601). Now tries
  GNU then busybox `-D`, and warns+skips on parse failure (no silent fail-open).
- docs: runbook + architecture document the CoreDNS handling.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/architecture/automated-upgrades.md

@@ -309,7 +309,11 @@ each Job's pod and its drain target are always different nodes.
   ConfigMap, and a `template` ConfigMap into each Job pod.
 - **Per-node script**: `infra/scripts/update_k8s.sh`. Caller passes
   `--role master|worker --release X.Y.Z`. Piped via SSH into each node by
-  upgrade-step.sh.
+  upgrade-step.sh. The master path runs `kubeadm upgrade apply` with
+  `--ignore-preflight-errors=CoreDNSMigration,CoreDNSUnsupportedPlugins
+  --skip-phases=addon/coredns` so kubeadm never touches CoreDNS (custom Corefile
+  + separately-tracked image; CoreDNS is pinned off Keel via `keel.sh/policy=never`).
+  See the runbook's "CoreDNS is NOT upgraded by kubeadm here".
 - **Four Upgrade Gates alerts**:
   - `K8sVersionSkew` — kubelet/apiserver `gitVersion` count >1 for 30m. Catches a half-done rollout.
   - `EtcdPreUpgradeSnapshotMissing` — `k8s_upgrade_in_flight==1 && k8s_upgrade_snapshot_taken==0` for 10m. Catches preflight failing silently.

docs/runbooks/k8s-version-upgrade.md

@@ -118,6 +118,26 @@ Pushed by upgrade-step.sh during phase execution; observed by the
 - **`K8sUpgradeChainJobFailed`** — `kube_job_status_failed{namespace="k8s-upgrade",job_name=~"k8s-upgrade-.*",reason=~"BackoffLimitExceeded|DeadlineExceeded"} > 0` for 15m (warning). Catches a phase Job that **terminally failed before `k8s_upgrade_in_flight` was set** — the preflight gates exit pre-metric, so the two `in_flight`-based alerts above are blind to a failed preflight (this is what hid the 5-day 1.34.9 wedge on 2026-06-12). Reason-scoped to terminal job conditions so a retry-success doesn't false-positive (a bare failed-pod-count would otherwise also block kured for the Job's 7d TTL).
 - All four alerts ALSO block kured (same `--prometheus-url` halt-on-alert mechanism) so the OS-reboot pipeline can't run on top of a half-done version upgrade.
 
+### CoreDNS is NOT upgraded by kubeadm here
+
+CoreDNS runs a **custom split-horizon Corefile** (owned by the technitium stack)
+and its image is tracked separately — it must NOT be touched by kubeadm. The
+master `kubeadm upgrade apply` therefore runs with
+`--ignore-preflight-errors=CoreDNSMigration,CoreDNSUnsupportedPlugins
+--skip-phases=addon/coredns` (in `scripts/update_k8s.sh`), so kubeadm upgrades
+the control plane but leaves CoreDNS 100% untouched (image + Corefile). Without
+the `--skip-phases`, forcing past the preflight makes kubeadm overwrite the
+Corefile with its default and downgrade the image (verified via
+`kubeadm upgrade apply --dry-run`).
+
+**Keep CoreDNS off Keel.** On 2026-06-12 Keel had auto-bumped CoreDNS
+v1.12.1 → v1.12.4 (kube-system out-of-band annotation from the 2026-05-26 Keel
+cascade), and 1.12.4 is ahead of kubeadm 1.34.9's corefile-migration table —
+which is what blocked the 1.34.9 upgrade. CoreDNS is now `keel.sh/policy=never`
+(`kubectl -n kube-system annotate deploy/coredns keel.sh/policy=never`). If a
+future kubeadm minor ships a CoreDNS that DOES know the running version, drop the
+`--skip-phases` for that run to let kubeadm re-take ownership.
+
 ### Vault secrets
 
 - `secret/k8s-upgrade/ssh_key` — ed25519 PRIVATE key, used by Jobs to SSH `wizard@<node>`

scripts/update_k8s.sh

@@ -98,7 +98,20 @@ if [[ "$ROLE" == "master" ]]; then
     # right version (which is the only case where this timeout fires).
     attempt=1
     extra_flags=""
-    while ! sudo kubeadm upgrade apply "v$RELEASE" -y $extra_flags; do
+    # CoreDNS is managed OUTSIDE kubeadm on this cluster: the Corefile is a
+    # custom split-horizon config owned by the technitium stack, and the image
+    # is intentionally tracked separately. kubeadm's bundled corefile-migration
+    # library rejects CoreDNS versions it doesn't know (e.g. 1.12.4 -> "start
+    # version not supported"), which HARD-FAILS `upgrade apply` at preflight.
+    # Forcing past preflight with --ignore alone is NOT enough — kubeadm would
+    # then overwrite our custom Corefile with its default AND downgrade the
+    # image (verified via `kubeadm upgrade apply --dry-run`, 2026-06-17). So we
+    # also skip the coredns addon phase entirely: kubeadm leaves CoreDNS 100%
+    # untouched and only upgrades the control-plane components. (Root fix: keep
+    # CoreDNS off Keel — keel.sh/policy=never — so it stops drifting ahead of
+    # kubeadm's migration table.)
+    coredns_flags="--ignore-preflight-errors=CoreDNSMigration,CoreDNSUnsupportedPlugins --skip-phases=addon/coredns"
+    while ! sudo kubeadm upgrade apply "v$RELEASE" -y $coredns_flags $extra_flags; do
         if (( attempt >= 3 )); then
             echo "ERROR: kubeadm upgrade apply failed after 3 attempts" >&2
             exit 1

stacks/k8s-version-upgrade/scripts/upgrade-step.sh

@@ -306,11 +306,19 @@ phase_preflight() {
   # reboot for an hour. 10min is sufficient for kubelet/control-plane to
   # stabilise; the kured-sentinel-gate DaemonSet enforces the broader
   # 24h-between-cluster-reboots invariant.
-  local recent=0
+  local recent=0 now_ep ts_ep
+  now_ep=$(date -u +%s)
   while IFS= read -r ts; do
     [ -z "$ts" ] && continue
-    local diff=$(( $(date +%s) - $(date -d "$ts" +%s) ))
-    if [ "$diff" -lt 600 ]; then recent=1; break; fi
+    # Portable ISO8601(UTC) -> epoch. GNU `date -d` parses ISO8601 directly;
+    # busybox `date` (the ghcr claude-agent-service base) does NOT and needs an
+    # explicit -D format. Before 2026-06-17 the bare `date -d "$ts"` silently
+    # failed on busybox, making this whole settle-window check a no-op. On
+    # parse failure, warn + skip the node (never silently treat it as quiet).
+    ts_ep=$(date -u -d "$ts" +%s 2>/dev/null || true)
+    if [ -z "$ts_ep" ]; then ts_ep=$(date -u -D '%Y-%m-%dT%H:%M:%SZ' -d "$ts" +%s 2>/dev/null || true); fi
+    if [ -z "$ts_ep" ]; then echo "WARN quiet-baseline: cannot parse Ready ts '$ts' (date impl?); skipping"; continue; fi
+    if [ "$(( now_ep - ts_ep ))" -lt 600 ]; then recent=1; break; fi
   done < <($KUBECTL get nodes -o jsonpath='{range .items[*]}{range .status.conditions[?(@.type=="Ready")]}{.lastTransitionTime}{"\n"}{end}{end}')
   if [ "$recent" -eq 1 ]; then
     slack "ABORT preflight — node transitioned Ready <10min ago (settle window)"