chore(ci): migrate all jobs to self-hosted macOS arm64 runner

[HongmingWang-Rabbit]

Summary

Migrates every CI job in this repo from GitHub-hosted ubuntu-latest to a self-hosted Apple-silicon runner ([self-hosted, macos, arm64], registered at the Molecule-AI org level) to sidestep GitHub-hosted minute rate limits.

What changed

.github/workflows/ci.yml — all 7 jobs migrated:

Job	Adaptation
platform-build	runs-on swap only (setup-go has native darwin-arm64)
canvas-build	runs-on swap only
mcp-server-build	runs-on swap only
e2e-api	Non-trivial: replaced services: postgres/redis with inline docker run startup steps (GHA services: is Linux-only on self-hosted runners). Ports remapped to 15432/16379 to avoid collision with whatever the host exposes on the standard ports. Containers named molecule-ci-postgres / molecule-ci-redis, torn down in if: always(). Postgres readiness still gated on pg_isready via docker exec.
shellcheck	Non-trivial: ludeeus/action-shellcheck is a Docker action (Linux-only). Replaced with a native shellcheck --severity=warning invocation against find tests/e2e -name '*.sh'.
canvas-deploy-reminder	runs-on swap only (uses gh api, pre-installed on runner)
python-lint	runs-on swap only

.github/workflows/publish-platform-image.yml — platform image publisher:

• Added docker/setup-qemu-action@v3 with platforms: linux/amd64.
• Added platforms: linux/amd64 to both docker/build-push-action@v5 calls. The self-hosted runner is arm64 but Fly tenant machines pull linux/amd64, so QEMU-emulated cross-arch builds are required. GHA cache-from/cache-to behavior unchanged.
• ⚠️ Expect slower builds — amd64 emulation under QEMU on an 8GB M1 is measurably slower than a native ubuntu-latest amd64 build. Monitor the first run closely.

Runner host prereqs (already set up on hongming-m1-mini)

• Docker Desktop installed + running (needed by e2e-api and publish-platform-image)
• docker symlinked into /opt/homebrew/bin
• shellcheck v0.11.0 installed into /opt/homebrew/bin
• gh, Go, Node, Python all on PATH; setup-* actions handle version pinning per job

Risks & things to watch on the first run

1. Docker Desktop must be running — if it crashes/quits, e2e-api and publish-platform-image fail. No graceful fallback.
2. QEMU cross-arch image builds are slow + RAM-heavy on the 8GB M1. If publish-platform-image OOMs, the mitigation is to split Dockerfile builds into multi-stage with a native Go cross-compile (no emulation needed for the Go build stage) — can be a follow-up PR.
3. Self-hosted runner has persistent filesystem state. Cached node_modules, Go build cache, etc. will grow across runs. Add a workspace-cleanup step or monitor _work/ disk usage.
4. Concurrency — this one runner processes jobs serially. Multiple PRs will queue.
5. Port collisions — CI postgres/redis use 15432/16379. If something else on the host takes those later, e2e-api will fail loudly.
6. services: removal means migrations check uses the known container name ($PG_CONTAINER env) instead of docker ps --filter ancestor=postgres:16, which is actually more reliable.

Test plan

• Watch the first run on this PR end-to-end — every job must turn green
• Confirm e2e-api postgres/redis containers are cleaned up after both success and failure runs
• Confirm publish-platform-image pushes a working linux/amd64 image to GHCR + Fly registry (manual workflow_dispatch on main after merge is a safer way to validate than waiting for the next platform-touching merge)
• Watch runner disk usage after a few runs — add cleanup step if _work/ grows unbounded