Symptom

PR #5476 sat "CI pending" for ~30 minutes after every workflow had finished. All 25 checks were green, but the required gate commit status was stuck at pending — "Waiting for required PR gates: unit", blocking merge indefinitely.

Root cause (evidenced from run logs)

Deploy Gate (.github/workflows/deploy-gate.yml) triggers on workflow_run.completed for Test / Typecheck / Lint Code / Security Audit, polls commits/<sha>/check-runs exactly once, and posts the gate status. Two failure modes combine:

1. Stale read, no retry. On PR feat(agent-readiness): SDKs in the agent view + homepage rel=alternate pointer (orank round 2) #5476 (head b33988c70), the unit check-run completed at 08:11:12Z. The final gate run evaluated at 08:12:19Z — 67s later — and the check-runs API still returned unit=pending (eventual consistency). Log from run 29990493776:

   changes=success docs-stats=success unit=pending ... security-audit=success
   pending=unit
   

   It posted gate=pending at 08:12:20Z. Since that was the LAST workflow-completion event for the SHA, nothing ever re-evaluated: the stale pending is permanent.
2. Last-writer-wins across concurrent gate runs. Each of the four watched workflows fires its own gate run; a run triggered early but scheduled late can overwrite a later run's correct success with pending. Same terminal state, different trigger.

Recovery (works today)

gh run rerun <latest-gate-run-id> — the rerun re-polls (now consistent) and posts the true result. That's how #5476 was unstuck (gate flipped to success at 08:39:52Z).

Proposed fix

In the gate step, when pending is non-empty and failed is empty (i.e. nothing has actually failed — we're only waiting), retry the check-runs poll 3–4 times with a 30s sleep before posting pending. That bounds the stale-read window at ~2 minutes and turns the terminal-stale case into a self-healing one. Optionally guard last-writer-wins by skipping the status POST when the existing gate status for the SHA is already success with a newer updated_at evaluation basis.

Evidence

• Stuck status: gate | pending | 2026-07-23T08:12:20Z vs unit | completed | success | 08:11:12Z on b33988c70
• Gate run that posted it: https://github.com/koala73/worldmonitor/actions/runs/29990493776
• Recovery rerun: same run ID, re-run at ~08:39Z → success | All required PR gates passed