CI: push images to in-cluster registry to bypass Traefik

The TeX Live layer (~3.5 GB) failed to push to registry.alocoq.fr:
Traefik severed the upload mid-stream ("client disconnected during blob
PUT ... unexpected EOF"), buildkit retried at the wrong offset, and the
registry returned "blob upload invalid".

Push to the in-cluster registry Service (registry.git.svc.cluster.local:5000)
instead, so the upload never traverses Traefik. Changes:
- buildctl outputs use registry.insecure=true (registry is plain HTTP)
- add a verso-buildkitd-config ConfigMap with buildkitd.toml marking the
  registry http/insecure, so the second build can pull the base image back
- the verso Deployment and rolling update reference the in-cluster image

NOTE: the cluster nodes' containerd must also treat
registry.git.svc.cluster.local:5000 as an insecure registry, otherwise
the kubelet image pull for the test deployment will fail. That is node-
level config outside this repo.

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

.gitea/workflows/deploy-verso.yml

@@ -19,6 +19,21 @@ jobs:
         run: |
           kubectl -n ci delete job verso-buildkit --ignore-not-found=true --wait=true
 
+          # buildkitd config: mark the in-cluster registry as http (insecure)
+          # so the second build can resolve/pull the base image we just pushed.
+          cat <<'EOF' | kubectl apply -f -
+          apiVersion: v1
+          kind: ConfigMap
+          metadata:
+            name: verso-buildkitd-config
+            namespace: ci
+          data:
+            buildkitd.toml: |
+              [registry."registry.git.svc.cluster.local:5000"]
+                http = true
+                insecure = true
+          EOF
+
           cat <<'EOF' | kubectl apply -f -
           apiVersion: batch/v1
           kind: Job
@@ -52,27 +67,40 @@ jobs:
                       - |
                         set -eux
 
+                        # Push to the in-cluster registry (plain HTTP) to bypass
+                        # the Traefik ingress, whose read timeout was killing the
+                        # multi-GB TeX Live layer upload mid-stream. The base
+                        # image is pulled back in for the second build, so the
+                        # registry must be marked insecure for both push and pull
+                        # (buildkitd.toml handles the pull/resolve side).
+                        REG=registry.git.svc.cluster.local:5000
+
                         buildctl-daemonless.sh build \
                           --frontend=dockerfile.v0 \
                           --local context=/workspace/repo \
                           --local dockerfile=/workspace/repo/server-ce \
                           --opt filename=Dockerfile-base \
-                          --output type=image,name=registry.alocoq.fr/verso-base:latest,push=true
+                          --output type=image,name=$REG/verso-base:latest,push=true,registry.insecure=true
 
                         buildctl-daemonless.sh build \
                           --frontend=dockerfile.v0 \
                           --local context=/workspace/repo \
                           --local dockerfile=/workspace/repo/server-ce \
                           --opt filename=Dockerfile \
-                          --opt build-arg:OVERLEAF_BASE_TAG=registry.alocoq.fr/verso-base:latest \
-                          --output type=image,name=registry.alocoq.fr/verso:latest,push=true
+                          --opt build-arg:OVERLEAF_BASE_TAG=$REG/verso-base:latest \
+                          --output type=image,name=$REG/verso:latest,push=true,registry.insecure=true
                     volumeMounts:
                       - name: workspace
                         mountPath: /workspace
+                      - name: buildkitd-config
+                        mountPath: /etc/buildkit
 
                 volumes:
                   - name: workspace
                     emptyDir: {}
+                  - name: buildkitd-config
+                    configMap:
+                      name: verso-buildkitd-config
           EOF
 
       - name: Wait for build
@@ -212,7 +240,7 @@ jobs:
               spec:
                 containers:
                   - name: verso
-                    image: registry.alocoq.fr/verso:latest
+                    image: registry.git.svc.cluster.local:5000/verso:latest
                     ports:
                       - containerPort: 80
                     env:
@@ -244,7 +272,7 @@ jobs:
       - name: Deploy Verso image
         run: |
           kubectl -n test set image deployment/verso \
-            verso=registry.alocoq.fr/verso:latest
+            verso=registry.git.svc.cluster.local:5000/verso:latest
 
           kubectl -n test rollout restart deployment/verso
           kubectl -n test rollout status deployment/verso --timeout=300s