fix(queue): retry a transient GitHub mergeable_state read before deferring to the next sweep

[JSONbored]

Summary

Live incident: metagraphed#8037 (and others) sat unmerged for ~6 minutes after being approved, gate-passing, and fully-autonomous-merge-configured (autonomy.merge: auto, requireApprovals: 0). GitHub computes mergeable_state asynchronously after a push/review and can return "unknown" (still computing) even on the forced live re-fetch taken moments after posting the approving review — the disposition correctly refused to merge into an unconfirmed state, but simply deferred to the next scheduled regate sweep (~6 minutes later) instead of retrying. That multi-minute window is exactly where an overlapping sibling PR can land first and base-conflict the original PR out from under it — metagraphed#8037's own review panel had already flagged 2 such overlapping siblings (#8040, #8036).

refreshLiveMergeState now retries a short, bounded number of times (2) specifically on "unknown" — "dirty"/"blocked"/"behind" are real, stable, non-computing states and are never retried — before falling through to the same defer-to-sweep behavior as before. Mirrors the identical GitHub-lag pattern already used for diff/files computation (fetchAndStorePullRequestFilesForReview).

Validation

• New regression tests in ci-resolution.test.ts: retries once and resolves to "clean" within the same pass; gives up after the retry cap and still returns "unknown" (unchanged fallback behavior); does not retry a real "dirty" state; does not retry an outright fetch failure.
• All 4 existing test files that mock fetchLivePullRequestMergeState directly (not through refreshLiveMergeState) are unaffected — confirmed by running them.
• npm run typecheck clean; npm run test:coverage (full, unsharded): 21,486 passed / 0 failed.
• Test-only delay override via the shared test/helpers/vitest-setup.ts default (from perf(test): eliminate SQLite fsync and real rate-limit backoff wait in tests #8553), so the retry's own logic is exercised without adding real wall-clock time to the suite.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[codecov]

Codecov Report

❌ Patch coverage is 88.88889% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 92.50%. Comparing base (12aa3e6) to head (5a42c5a).
⚠️ Report is 4 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/queue/ci-resolution.ts	88.88%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #8560      +/-   ##
==========================================
- Coverage   92.51%   92.50%   -0.01%     
==========================================
  Files         795      795              
  Lines       79613    79622       +9     
  Branches    24056    24058       +2     
==========================================
+ Hits        73653    73657       +4     
  Misses       4800     4800              
- Partials     1160     1165       +5     

Flag	Coverage Δ
backend	93.67% <88.88%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/queue/ci-resolution.ts	97.59% <88.88%> (-1.06%)	⬇️

... and 1 file with indirect coverage changes