spec(queen): a conflicting bee PR is updated or redone, never stranded #7571

Context

Measured 2026-10-07 15:51Z (gh pr list --state open --json mergeable,headRefName): 94 of 213 open pull requests conflict with master, 67 of them on queen- branches, the oldest opened 2026-09-21. GitHub runs no required check on a conflicting head, so the reviewer bee never judges it and the merger never merges it. No bee updates its own branch, so every master merge that touches a bee's files strands that pull request for good. tools/queen/pusher.py already counts conflicted_bee_prs; nothing acts on the count.

Boundary

specs/queen/review_valve.t27, gen/c/queen/review_valve.c (generated), one docs/now/ entry. Wiring into the publisher and the supervisor is a separate issue.

Scenarios

• A bee PR conflicts and fewer than 50 Actions runs are queued: GitHub's update-branch merges master into the head (no force-push).
• Same, with the queue over 50: wait for the next tick, at most 48 hours.
• update-branch answers 422 (a real conflict), or the head conflicted again after 2 updates, or it waited 48 hours: close the PR with the reason and release the dispatch, so a fresh bee redoes the issue from current master. This is one of MAX_RELEASES; once those are spent, close.
• A human branch or a clean PR: not this valve's.

Requirements

• conflict_step(queen_branch, conflicting, update_failed, updates, hours, releases, queued_runs) -> u8 returning C_NONE / C_WAIT / C_UPDATE / C_REDO / C_CLOSE.
• Constants CONFLICT_UPDATES = 2, UPDATE_QUEUE_CEILING = 50, CONFLICT_WAIT_HOURS = 48.
• Step codes are pinned by an invariant, and the wait ends before the policy's 48-hour hold.

Success criteria

• t27c parse specs/queen/review_valve.t27 exits 0
• t27c typecheck specs/queen/review_valve.t27 exits 0
• t27c test-report specs/queen/review_valve.t27 reports vacuous passes 0