test-merge-queue

Lab for the test-app merge train. Nothing here is production code.

What is being tested

test-app merges are already strictly serialised by strict_required_status_checks_policy, but nothing decides whose turn is next. A gate run takes ~8 minutes and merges arrive every ~17, so a pull request can go green and already be stale. One pull request lost that race nine times in 4h20m on a five-file change.

The fix under test: a Queue position commit status that reserves one turn at a time.

Layout

File	Purpose
.github/workflows/mock-gates.yml	Five fake gates named after the real test-app checks, on ubuntu-latest, with a deliberate fast/slow spread (10s–45s) so the starvation race is reproducible. Put MOCK_FAIL=frontend in a PR body to fail one on purpose.
.github/workflows/pr-merge-train.yml	The mechanism under test. Elects one head, posts Queue position on every open PR, arms native auto-merge, and updates only the head's branch. Never merges directly.

Knobs

• Repo variable MERGE_TRAIN_EJECT_MINUTES — shortens the 30-minute eject so it can be tested.
• workflow_dispatch input simulate_crash — forces the sweep to throw, to prove it fails open.

Differences from test-app

• ubuntu-latest instead of the self-hosted runner.
• GITHUB_TOKEN instead of a GitHub App token. test-app needs the App token because a branch push made with GITHUB_TOKEN does not re-trigger CI — this lab measures what happens without it.
• Public repo, because rulesets are unavailable on a private repo under a free personal account.