ci: build against a dind sidecar instead of the host socket

The host docker socket wasn't mounted consistently across act_runners,
so which runner picked up the push was the difference between a clean
build and "unable to connect to docker." Switch to a dockerd service
pinned to the same bridge network as the job; the CLI now talks to
it over TCP via DOCKER_HOST=tcp://dockerd:2375. Every runner has a
daemon, every time — no more re-runs for the scheduling lottery.

.forgejo/workflows/build.yaml

@@ -18,6 +18,22 @@ jobs:
     name: Build and Push Docker Image
     runs-on: ubuntu-22.04
 
+    # Runs a dockerd sidecar on the same bridge network as the job
+    # container so the build doesn't depend on the host's docker socket
+    # being mounted. Some act_runners don't mount it and there was no
+    # way to tell at schedule time which runner would win — leading to
+    # random "unable to connect to docker" failures that went away on a
+    # re-run. With dind every runner has a daemon, every time.
+    services:
+      dockerd:
+        image: docker:27-dind
+        options: --privileged
+        env:
+          DOCKER_TLS_CERTDIR: ""
+
+    env:
+      DOCKER_HOST: tcp://dockerd:2375
+
     steps:
       - name: Checkout code
         uses: https://github.com/actions/checkout@v4
@@ -26,8 +42,8 @@ jobs:
         run: |
           # Dodge apt entirely — the runner image's bookworm repos have been
           # failing GPG verification with "At least one invalid signature".
-          # The static Docker client binary talks to the host's daemon socket
+          # The static Docker client binary talks to the dind sidecar via
-          # that act_runner mounts into the job container. buildx is needed
+          # the DOCKER_HOST env var set at the job level. buildx is needed
           # separately because the Dockerfile uses `--mount=type=cache`,
           # which is a BuildKit-only feature.
           curl -fsSL -o /tmp/docker.tgz \
@@ -44,28 +60,22 @@ jobs:
           docker --version
           docker buildx version
 
-      - name: Wait for Docker daemon
+      - name: Wait for dockerd sidecar
         run: |
-          # act_runner mounts the host docker socket into the job, but
+          # dind takes ~5-15s to finish booting inside the service
-          # the daemon side sometimes takes a moment to answer. Poll
+          # container; poll until the TCP endpoint answers.
-          # `docker info` until it responds so the login/build steps
+          echo "Using DOCKER_HOST=${DOCKER_HOST}"
-          # below don't race the socket and fail with "unable to
-          # connect to docker".
-          sock="${DOCKER_HOST:-unix:///var/run/docker.sock}"
-          echo "Using DOCKER_HOST=${sock}"
 
-          for i in $(seq 1 30); do
+          for i in $(seq 1 60); do
             if docker info >/dev/null 2>&1; then
-              echo "Docker daemon reachable after ${i}s"
+              echo "dockerd reachable after ${i}s"
               exit 0
             fi
-            echo "Waiting for Docker daemon... ($i/30)"
+            echo "Waiting for dockerd... ($i/60)"
             sleep 1
           done
 
-          echo "Docker daemon never became reachable — diagnostics:" >&2
+          echo "dockerd never became reachable — diagnostics:" >&2
-          echo "--- ls -l /var/run/docker.sock ---" >&2
-          ls -l /var/run/docker.sock 2>&1 || true
           echo "--- env | grep -i docker ---" >&2
           env | grep -i docker 2>&1 || true
           echo "--- docker info (verbose) ---" >&2