fix(build): version-match the containerd pause image to kubeadm (C4)

.github/workflows/ci.yml

@@ -157,6 +157,20 @@ jobs:
           esac
           echo "OK: image bundles the k8s ${{ matrix.minor }} toolchain"
 
+      - name: Verify containerd sandbox (pause) image is version-matched (C4)
+        run: |
+          set -euo pipefail
+          img="kairos-kubeadm:ci-${{ matrix.minor }}"
+          # The build pins containerd's sandbox_image to the pause kubeadm expects;
+          # assert they match so a stale/hardcoded pause tag can't regress (C4).
+          want="$(docker run --rm --entrypoint /usr/bin/kubeadm "$img" \
+            config images list --image-repository registry.k8s.io | grep -E '/pause:[0-9]')"
+          got="$(docker run --rm --entrypoint sh "$img" -c \
+            "grep -oE 'registry.k8s.io/pause:[0-9.]+' /etc/containerd/config.toml")"
+          echo "kubeadm pause: ${want}; containerd sandbox_image: ${got}"
+          [ "$got" = "$want" ] || { echo "pause mismatch (C4): containerd ${got} != kubeadm ${want}" >&2; exit 1; }
+          echo "OK: sandbox_image is version-matched to kubeadm"
+
       - name: Verify kubeadm exposes the upgrade-path surface (ADR-12)
         run: |
           set -euo pipefail

Dockerfile

@@ -260,6 +260,21 @@ COPY systemd/kubelet.service.d/10-kubeadm.conf /etc/systemd/system/kubelet.servi
 COPY sysctl/k8s.conf                        /etc/sysctl.d/k8s.conf
 COPY modules-load/k8s.conf                  /etc/modules-load.d/k8s.conf
 
+# Pin containerd's pod-sandbox (pause) image to the EXACT version the bundled
+# kubeadm expects for this Kubernetes minor, instead of a hardcoded tag. kubeadm
+# bumps the pause version per release (e.g. 3.10.1 in 1.34/1.35 -> 3.10.2 in 1.36);
+# a stale tag means containerd pulls a different pause than kubeadm pre-pulled
+# (duplicate image / drift -- pitfall C4). Resolved from kubeadm at build time so
+# it always matches the bundled toolchain. kubeadm is static and runs here on musl.
+RUN set -eux; \
+    pause="$(/usr/bin/kubeadm config images list \
+      --kubernetes-version "${KUBERNETES_VERSION}" \
+      --image-repository registry.k8s.io | grep -E '/pause:[0-9]')"; \
+    test -n "${pause}"; \
+    sed -i "s#^\([[:space:]]*sandbox_image[[:space:]]*=\).*#\1 \"${pause}\"#" /etc/containerd/config.toml; \
+    grep -q "sandbox_image = \"${pause}\"" /etc/containerd/config.toml; \
+    echo "pinned containerd sandbox_image to ${pause}"
+
 # --- Boot-time setup: enable services; modules and sysctls load via /etc -----
 RUN systemctl enable containerd.service kubelet.service
 

containerd/config.toml

@@ -1,12 +1,15 @@
 # containerd configuration for kubeadm clusters built with provider-kubernetes.
-# - sandbox image pinned to registry.k8s.io (NOT the dead k8s.gcr.io).
+# - sandbox_image uses registry.k8s.io (NOT the dead k8s.gcr.io). The pause TAG
+#   below is a placeholder: the image build overwrites it with the exact pause
+#   version the bundled kubeadm expects for the target Kubernetes minor (see the
+#   Dockerfile), so it is always version-matched rather than a stale hardcoded tag.
 # - SystemdCgroup = true, matching the kubelet's default cgroup driver
 #   ("systemd"); see https://kubernetes.io/docs/setup/production-environment/container-runtimes/.
 version = 2
 root = "/var/lib/containerd"
 
 [plugins."io.containerd.grpc.v1.cri"]
-  sandbox_image = "registry.k8s.io/pause:3.10"
+  sandbox_image = "registry.k8s.io/pause:3.10.1"
 
   [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc]
     runtime_type = "io.containerd.runc.v2"