ci: workflow-level 3-attempt retry around buildx build --push

Belt-and-braces against transient registry-1.docker.io blips (rate
limits, brief 5xx, CDN flap). Replaces docker/build-push-action@v7 with
a shell: bash step that runs docker buildx build --push in a for-loop
with backoff (15s, 30s).

Does NOT mask deterministic failures: a true regression (e.g. the
cache-export 400 we hit 2026-05-23..28) fails all 3 attempts
identically and the job still fails by design. Orthogonal layer to
both cache-export disablement and the ci-release-watcher skill's
transient-rerun heuristic.

No image-side change.

.gitea/workflows/docker-publish.yml

@@ -112,15 +112,22 @@ jobs:
           echo "pi_version=${PI_VERSION}" >> "$GITHUB_OUTPUT"
           echo "Resolved PI_VERSION=${PI_VERSION} from tag ${TAG}"
 
-      - name: Build and push (amd64 + arm64)
+      - name: Build and push (amd64 + arm64) — with retry
-        uses: docker/build-push-action@v7
+        shell: bash
-        with:
+        env:
-          context: .
+          PI_VERSION: ${{ steps.resolve.outputs.pi_version }}
-          platforms: linux/amd64,linux/arm64
+          TAGS: ${{ steps.tags.outputs.tags }}
-          push: true
+        run: |
-          tags: ${{ steps.tags.outputs.tags }}
+          set -euo pipefail
-          build-args: |
+          # Convert newline-delimited TAGS env var (build-push-action's native
-            PI_VERSION=${{ steps.resolve.outputs.pi_version }}
+          # format from the `Compute tags` step) into a bash array of -t flags.
+          TAG_FLAGS=()
+          while IFS= read -r t; do [[ -n "$t" ]] && TAG_FLAGS+=( -t "$t" ); done <<< "${TAGS}"
+          # 3-attempt retry around `docker buildx build --push` for transient
+          # registry-1.docker.io blips (rate limits, CDN flap, brief 5xx).
+          # Does NOT mask deterministic failures: a true regression (e.g. the
+          # cache-export 400 hit 2026-05-23..28) will fail all 3 attempts
+          # identically and the job still fails — by design.
           # Registry cache disabled: buildkit's mode=max cache-export to
           # registry-1.docker.io reproducibly returns HTTP 400 on resumable-
           # upload PUT (Hub-CDN protocol mismatch with buildx 0.34.x, surfaced
@@ -130,6 +137,25 @@ jobs:
           # root-cause analysis. Re-enable when buildkit upstream resolves.
           # Single-stage Dockerfile + tiny diff (npm install pi only) means
           # build is fast even without cache (~30-60s).
+          for attempt in 1 2 3; do
+            echo "==> Build+push attempt ${attempt}/3"
+            if docker buildx build \
+                --platform linux/amd64,linux/arm64 \
+                --push \
+                --build-arg "PI_VERSION=${PI_VERSION}" \
+                "${TAG_FLAGS[@]}" \
+                .; then
+              echo "==> Attempt ${attempt} succeeded"
+              exit 0
+            fi
+            if [[ "${attempt}" -lt 3 ]]; then
+              backoff=$(( attempt * 15 ))
+              echo "==> Attempt ${attempt} failed, sleeping ${backoff}s before retry"
+              sleep "${backoff}"
+            fi
+          done
+          echo "==> All 3 build+push attempts failed"
+          exit 1
 
   update-description:
     needs: publish

CHANGELOG.md

@@ -8,7 +8,13 @@ Tags follow the pi npm version: `v{pi_version}[letter]` — bare tag for the fir
 
 ## Unreleased
 
-_(no changes since v0.76.0)_
+### CI: workflow-level retry around `docker buildx build --push`
+
+The single push step in `.gitea/workflows/docker-publish.yml` is now wrapped in a 3-attempt retry loop with backoff (15s, 30s) as belt-and-braces against transient `registry-1.docker.io` blips (rate limits, brief 5xx, CDN flap). Replaces the `docker/build-push-action@v7` invocation with a `shell: bash` step that runs `docker buildx build --push` directly so the loop is visible and tweakable.
+
+Does **not** mask deterministic failures: a true regression (e.g. the cache-export 400 we hit 2026-05-23..28) will fail all 3 attempts identically and the job still fails by design. Only intended to absorb genuinely transient failures that single-attempt CI was vulnerable to.
+
+No image-side change.
 
 ## v0.76.0 — 2026-05-28