Closing-target claim guard reads a yielded claim as the newest claim and reds the winning branch #25791

Filing gate: ① a defect measured on a live PR.
Reader: domain:skills seat, the next PR on scripts/check-closing-target-claim.mjs and the claim reader it imports (claimGovernance / h46ClaimNamesBranch in scripts/pm/check-half-states.mjs).
Dedupe (open AND closed, the 300 newest issues by title, local grep for yield / closing-target / claim guard / 让行 / already claimed): 0 hits. /search is refused by the egress proxy, so the dedupe is title-only over that window.

What happened (PR #25780, card #25729, 2026-10-25)

• Seat 1 claimed the card at 09:47Z (comment 6396126096, Branch: claude/project-thread-biervl).
• Seat 2 claimed the same card at 09:54Z (6396193561, Branch: claude/issue-25729-adr-tier-row), read the thread again, and yielded at 10:01Z (6396249639, first line already claimed — yielding), exactly as execution-duties.md 〈认领〉 ③ prescribes. The branch it named was never created.
• The PR's "The card this PR closes must claim this branch" check failed (run 40183909709): the guard reads the NEWEST Claim: comment, which was seat 2's yielded one, and judged the card claimed on another branch.
• It went green only after seat 1 re-posted an identical claim (6396460178) and edited the PR body to re-run the guard.

Why it matters

The race re-read rule makes the loser yield in a comment, never by deleting its claim. So every lawful two-seat race on one card leaves the winner's PR red until a human or the seat notices and re-posts. That costs a cycle on exactly the cards where two seats were both live.

Wanted

A Claim: comment that is answered later on the same thread by a yield from the same session (the protocol's already claimed — yielding line, which names that session) is not the card's current claim. The guard skips it and reads the newest claim that is not yielded. A yield that names no session, or a later Claim: from the yielding session, keeps today's behaviour. If the yield spelling is not machine-readable enough to key on, the fix may instead define one fixed first line for the yield and teach that line in execution-duties.md 〈认领〉 ③ in the same PR.

Acceptance

• A self-test fixture reproduces this thread (claim A → claim B → B yields) and the guard passes for branch A and fails for branch B.
• The existing fixtures stay green; pnpm check:pm-skill-ratchet and check:pm-skill-id-lint stay green if skill text moves.
• Writes: PR only through the relay's pr_create; ⛔ GitHub MCP writes.

Session: session_014FZ3TJucS1L7BdMe5DAzXC (domain:skills#1, seat post #7623)