Summary

On 2026-07-21 a burst of 6 PRs from one contributor against a cap of 2 exposed two bugs in the per-contributor open-item cap (PRs #7284-#7289): only 2 of the 4 excess PRs were caught, and PRs merged that should have been closed.

Ordering (resource waste): the cap check only ran as the LAST step of the full review pipeline (miner detection, AI slop advisory, linked-issue satisfaction, the full AI code review, CI-completeness verification, public surface publish), so an over-cap PR paid for that entire pipeline before being caught. Confirmed via the real audit log: contributor_cap_ci_cancelled for the burst logged "cancelled 0 of 0 in-flight workflow run(s)" — CI had already finished by the time the close fired. Adds a cheap, CI-independent short-circuit on pull_request opened that closes an over-cap PR before any of that work runs. The existing end-of-pipeline check stays in place as defense-in-depth for state that changes mid-pipeline. A repo with no cap configured, an ignored/exempt author, or a lock-contended pass pays nothing extra — byte-identical to today.

Race (TOCTOU): cap membership was recomputed independently by every sibling PR's own pipeline pass, using whichever "confirmed still open" siblings existed at that exact moment — not one shared snapshot for the burst. A merge could therefore execute off a stale "under cap" conclusion after a sibling's later close made the author over cap. Adds a per-(repo, author) mutex (the existing atomic claim/release transient-lock primitive already used for the per-PR actuation lock) around the cap decision, plus a re-check under that same mutex immediately before a merge actually executes — redirecting to a "denied" outcome instead of merging when a fresh check confirms the author is now over cap, so the next natural re-evaluation plans a close off accurate state.

prReadyForReview's own CI-wait/staleness-cap/dirty-base internals are completely untouched by either fix.

Test plan

• Full regression sweep across every test file touching the modified functions (queue, queue-2/3/4, agent-actions, agent-action-executor, ci-completion-fork-resume, pr-panel-retrigger-pending-force-review) — 1010/1010 passing
• New regression tests: early-path close before the expensive pipeline runs, resource-waste guard (no cap configured → zero added cost), ignore-authors parity, new-account label preserved, lock-contended fail-open (both the early path and the pre-merge re-check), pre-merge re-check redirects a would-be merge to denied when now over cap, lock release on both success and denial paths, full-pipeline integration test proving an under-cap contributor still merges normally
• npm run test:coverage (full, unsharded) — all green, no regressions
• Coverage scoped to the diff: 100% of changed lines and branches covered across all four modified source files (verified via lcov DA/BRDA against the diff hunks); remaining pre-existing/defensive branch arms carry justified v8 ignore comments mirroring this file's own established convention
• npx tsc --noEmit clean
• npm audit --audit-level=moderate clean