ci(deploy): publish the stack over TLS when PUBLIC_DOMAIN is set (refs #175)

The in-cluster edge (#178) is off unless the chart is given a domain, so pass
one through from a repository variable. Unset, the deploy is exactly what it was
— NodePorts, and the portals reachable only over the SSH forwards.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.gitea/workflows/deploy.yaml

@@ -0,0 +1,126 @@
+name: Deploy to Talos
+
+# A merge to main ships the stack to the Talos cluster on the lab server
+# (docs/runbooks/kubernetes-talos.md §9). PR CI is the merge gate, so main is
+# green by construction — this workflow only deploys.
+on:
+  push:
+    branches: [main]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+# Queue deploys, never cancel one: a helm upgrade killed half-way leaves the
+# release in `pending-upgrade` and the next run has to be unwedged by hand.
+concurrency:
+  group: deploy-talos
+  cancel-in-progress: false
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    env:
+      # The Talos VM as seen from the Fedora host (libvirt guest IP), and the
+      # address a browser uses to reach the cluster. `localhost` is deliberate:
+      # the portals' PKCE needs a secure context, so they are reached over
+      # `kubectl port-forward` — runbook §5. Override with repo variables.
+      TALOS_VM_IP: ${{ vars.TALOS_VM_IP }}
+      TALOS_HOST: ${{ vars.TALOS_HOST }}
+      # Set it and the stack is published over TLS on <sub>.<domain> by the
+      # in-cluster edge (ADR-0035, runbook §10). Empty = NodePorts, as before.
+      PUBLIC_DOMAIN: ${{ vars.PUBLIC_DOMAIN }}
+      PUBLIC_EMAIL: ${{ vars.PUBLIC_EMAIL }}
+    steps:
+      - uses: https://github.com/actions/checkout@v4
+
+      # Pinned static binaries, the same URLs the Talos runbook §0 gives a
+      # developer and the same helm the `k8s` CI job uses — no action to vet.
+      - name: Install kubectl, helm and crane
+        run: |
+          set -euo pipefail
+          bin="$HOME/.local/bin"; mkdir -p "$bin"
+          curl -sSLo "$bin/kubectl" https://dl.k8s.io/release/v1.37.0/bin/linux/amd64/kubectl
+          curl -sSL https://get.helm.sh/helm-v3.16.4-linux-amd64.tar.gz | tar xz -O linux-amd64/helm > "$bin/helm"
+          curl -sSL https://github.com/google/go-containerregistry/releases/download/v0.20.2/go-containerregistry_Linux_x86_64.tar.gz | tar xz -O crane > "$bin/crane"
+          chmod +x "$bin"/{kubectl,helm,crane}
+          echo "$bin" >> "$GITHUB_PATH"
+
+      # The cluster's API and its registry are only reachable through the Fedora
+      # host, so forward both to the runner. 30141 is the openbaar portal, for
+      # the smoke at the end.
+      - name: Tunnel the Talos API + registry through the Fedora host
+        env:
+          SSH_KEY: ${{ secrets.TALOS_SSH_KEY }}
+        run: |
+          set -euo pipefail
+          : "${TALOS_VM_IP:=192.168.122.173}"
+          umask 077
+          printf '%s\n' "$SSH_KEY" > ~/.ssh_talos
+          ssh -i ~/.ssh_talos -o StrictHostKeyChecking=no -o IdentitiesOnly=yes \
+              -o ExitOnForwardFailure=yes -p 6667 -f -N \
+              -L 6443:$TALOS_VM_IP:6443 \
+              -L 30500:$TALOS_VM_IP:30500 \
+              -L 30141:$TALOS_VM_IP:30141 \
+              user@labs.respellion.tech
+
+      # The kubeconfig's server must be https://127.0.0.1:6443 — Talos puts
+      # 127.0.0.1 in the apiserver cert SANs, so TLS verification still holds
+      # through the tunnel.
+      - name: Write the kubeconfig
+        env:
+          KUBECONFIG_B64: ${{ secrets.TALOS_KUBECONFIG }}
+        run: |
+          set -euo pipefail
+          umask 077
+          base64 -d <<< "$KUBECONFIG_B64" > "$RUNNER_TEMP/kubeconfig"
+          echo "KUBECONFIG=$RUNNER_TEMP/kubeconfig" >> "$GITHUB_ENV"
+          kubectl --kubeconfig "$RUNNER_TEMP/kubeconfig" get nodes
+
+      # Idempotent; also makes a first deploy onto a bare cluster work. The
+      # registry's storage is an emptyDir, so a replaced pod loses the images —
+      # which the push in the next step puts back anyway.
+      - name: Ensure the in-cluster registry
+        run: make k8s-registry
+
+      # Push through the tunnel (localhost), pull from the node's own NodePort
+      # (the address in the Talos registry-mirror patch) — same registry, two
+      # names, so the two `make` calls get different K8S_REGISTRY values.
+      - name: Build and push the images
+        run: make k8s-images K8S_REGISTRY=localhost:30500
+
+      # k8s-reseed = seed configmaps + helm upgrade + re-run the bootstrap jobs.
+      # The jobs are idempotent, and deleting them first is what keeps a changed
+      # Job template from wedging the upgrade (`cannot patch … with kind Job`).
+      - name: Deploy the chart
+        run: |
+          set -euo pipefail
+          publish="${PUBLIC_DOMAIN:+--set public.domain=$PUBLIC_DOMAIN --set public.email=${PUBLIC_EMAIL:-}}"
+          make k8s-reseed \
+            TALOS_HOST=${TALOS_HOST:-localhost} \
+            K8S_REGISTRY=${TALOS_VM_IP:-192.168.122.173}:30500 \
+            K8S_SET="$publish"
+
+      # `dev` is a mutable tag and helm sees an unchanged pod template, so the
+      # new images only land on a restart (pullPolicy is already Always).
+      - name: Roll the services onto the new images
+        run: |
+          set -euo pipefail
+          svcs="acl domain bff event-subscriber projection-api self-service openbaar behandel beheer"
+          kubectl -n big rollout restart deploy $svcs
+          kubectl -n big rollout status --timeout=300s deploy $svcs
+
+      # Proves portal → Caddy → BFF → projection end to end. An empty register is
+      # a pass; a 502 or a timeout is not.
+      - name: Smoke the public register
+        run: curl -fsS --retry 10 --retry-delay 6 --retry-all-errors http://localhost:30141/openbaar/register
+
+      # Cluster-wide, not just `big`: the first thing that can fail is the registry
+      # in its own namespace, and a scheduling problem shows up in the events, not
+      # in `rollout status` — which only ever says "timed out waiting".
+      - name: Pods and events on failure
+        if: failure()
+        run: |
+          kubectl get pods -A -o wide || true
+          kubectl -n big get jobs || true
+          kubectl get events -A --sort-by=.lastTimestamp | tail -30 || true

docs/runbooks/ci.md

@@ -31,6 +31,10 @@ and CI cannot drift:
 > services by **container IP** (the runner can't reach published ports — see
 > [gitea-actions-gotchas.md §5/§6](gitea-actions-gotchas.md)).
 
+A second workflow, `.gitea/workflows/deploy.yaml`, deploys the stack to the Talos
+cluster on the lab server when a PR is merged to `main` — see
+[kubernetes-talos.md §9](kubernetes-talos.md) for its secrets and the SSH tunnel it needs.
+
 All `uses:` references are absolute, tag-pinned URLs (`https://github.com/actions/checkout@v4`,
 `https://github.com/actions/setup-dotnet@v4`) per CLAUDE.md §8.7 and §15 — Gitea
 Actions resolves them from GitHub.

docs/runbooks/kubernetes-talos.md

@@ -360,6 +360,97 @@ immutable, so `helm upgrade` is rejected with `cannot patch "…" with kind Job`
 | Pods `Evicted` / `OOMKilled` | the VM is too small (§0) |
 | A Job shows `BackoffLimitExceeded` | read it: `kubectl -n big logs job/<name>` |
 
+## 9. Deploying on merge to main
+
+`.gitea/workflows/deploy.yaml` runs the §3–§4 steps against the **lab server's** Talos VM
+every time a PR is squash-merged to `main` (and on demand via *Run workflow*). PR CI is the
+merge gate, so the workflow deploys without re-running the checks.
+
+**Prerequisite: the VM must have been installed with the §1 patch.** A stock Talos config
+gives you a node that still carries the control-plane taint and knows nothing about the
+plain-HTTP registry, and the deploy hits those in that order: the `registry` pod sits
+`Pending` until `rollout status` times out, and once that is fixed every repo image fails to
+pull. Two separate fixes:
+
+```bash
+# on the Fedora host — 1. let workloads onto the only node (§1)
+export KUBECONFIG=~/talos-kubeconfig-local
+kubectl taint node --all node-role.kubernetes.io/control-plane-
+
+# 2. trust the in-cluster registry over plain HTTP (§2)
+cat > /tmp/registry-patch.yaml <<'YAML'
+machine:
+  registries:
+    mirrors:
+      "<TALOS_VM_IP>:30500":
+        endpoints:
+          - http://<TALOS_VM_IP>:30500
+YAML
+talosctl -n <TALOS_VM_IP> -e <TALOS_VM_IP> patch mc --patch @/tmp/registry-patch.yaml
+```
+
+Keep those two apart. On Talos 1.14 a patch that also sets
+`cluster.allowSchedulingOnControlPlanes` is rejected with *".cluster.allowSchedulingOnControlPlanes
+is already set in v1alpha1 config"* — the field moved out of the v1alpha1 schema, the same way
+`machine.install` did (§1) — and the rejection takes the whole patch with it, so the mirror
+silently doesn't land either. `kubectl taint` is the documented way (§1); it is undone if the
+node ever re-registers, which is a reboot, not a deploy.
+
+The cluster's API and registry are not exposed publicly, so the job forwards them over the
+same SSH hop the Gitea-runner pipeline uses:
+
+```
+ssh -p 6667 user@labs.respellion.tech  -L 6443 -L 30500 -L 30141  →  <TALOS_VM_IP>
+```
+
+Consequences worth knowing:
+
+- Images are **pushed** to `localhost:30500` (the tunnel) and **pulled** by the node from
+  `<TALOS_VM_IP>:30500` (its own NodePort, the address in the Talos registry-mirror patch).
+  Same registry, two names — hence the two `K8S_REGISTRY` values in the workflow.
+- It calls `make k8s-reseed`, not `make k8s-up`: the bootstrap Jobs are idempotent, and
+  deleting them first is what stops a changed Job template from wedging `helm upgrade` (§7).
+- `dev` is a mutable tag, so a `rollout restart` of the nine repo deployments is what
+  actually puts the new images in the pods.
+- Deploys **queue** (`cancel-in-progress: false`): a helm upgrade killed half-way leaves the
+  release in `pending-upgrade`, which has to be unwedged by hand.
+
+Settings, all on the repository in Gitea:
+
+| Kind | Name | What |
+|---|---|---|
+| Secret | `TALOS_SSH_KEY` | private key for `user@labs.respellion.tech` (the Fedora host) |
+| Secret | `TALOS_KUBECONFIG` | base64 of the kubeconfig, **`server: https://127.0.0.1:6443`** — Talos puts `127.0.0.1` in the apiserver cert SANs, so TLS still verifies through the tunnel |
+| Variable | `TALOS_VM_IP` | the VM's libvirt address (default `192.168.122.173`) |
+| Variable | `TALOS_HOST` | the browser-facing host baked into Keycloak's issuer (default `localhost`, see §5) |
+
+The last step smokes `GET /openbaar/register` through the openbaar portal, which exercises
+portal → Caddy → BFF → projection. An empty register passes; a 502 does not.
+
+### Reaching the portals from a laptop
+
+The deployed portals are pinned to `http://localhost:30180` for Keycloak (§5), so a browser
+needs **all five** browser-facing ports on its own localhost — the portal alone is not
+enough, and a missing Keycloak shows up as `ERR_CONNECTION_REFUSED` on
+`/realms/*/.well-known/openid-configuration` followed by an opaque `ERROR Error: [object Object]`.
+`make k8s-portals` does this when kubectl can reach the cluster; through the lab server one
+SSH does it without a kubeconfig at all:
+
+```bash
+ssh -N -p 6667 \
+  -L 30140:<TALOS_VM_IP>:30140 -L 30141:<TALOS_VM_IP>:30141 \
+  -L 30142:<TALOS_VM_IP>:30142 -L 30143:<TALOS_VM_IP>:30143 \
+  -L 30180:<TALOS_VM_IP>:30180 \
+  user@labs.respellion.tech
+```
+
+Then the §5 table's URLs work as written. The admin UIs (OpenZaak, Flowable, …) need no
+forward — they are server-rendered, so the VM's address is fine.
+
+Not covered: the portals still need `make k8s-portals` (or an SSH forward) to be usable in a
+browser, because PKCE needs a secure context (§5). Giving the server a hostname + TLS is the
+upgrade path.
+
 ## What is not ported
 
 - **Observability** (Tempo, Prometheus, Grafana) is defined but disabled — those are built