fix(pm): tell a completed clause-② review from a gate that never ran

[claude]

Fixes #14155

check-clause2-carriers.mjs --pair answered C3 / exit 4 on a legitimately cleared clause-② pair. A Clause-②: yes declaration is history — it stays on the thread forever. The gate label is state — a completed review clears it from both carriers, by rule. So every clause-② pair that completes its review lands in exactly C3's trigger shape: declared yes, gate on neither carrier. Read literally, precondition ② of the landing check could only be satisfied in the window between the review PASS and the label clear — the wrong order — and each landed pair kept its C3 row on every later sweep.

What changed

C3 (and therefore --pair) now reads the needs:contract-review label event stream on both carriers and answers four distinguishable states:

events say	verdict	exit
never hung on either carrier	C3 fires — the fail-open, unchanged	4
hung then cleared on both, head unmoved since	completed state — clean	0
hung then cleared on both, head moved since	adverse: the re-hang-owed state	4
bound on one carrier only	adverse: one removal where a clear leaves two — the strip signature	4
stream or head commit unreadable	UNJUDGED, never clean	2

This mechanizes the recovery rule in references/contract-review.md — 「PASS + 无标 + head 未动 = 已清标非被剥;head 后移或无结论才重挂」 — minus its PASS conjunct, which stays human. Precondition ① is still a person reading the PASS comment on the card.

The three standing refusals are intact

• No verdict comment is read. The PASS half is deliberately not mechanized: [finding] The contract-review gate is being hung on ONE carrier — 34 of 36 lone clears had a PR that never carried it, which makes a strip undetectable #12409 measured a strict Contract review: PASS marker at 5 of 35 removals against 26 for an "any PASS token" reading — a check that can barely fail — and reading a verdict to clear a gate is the 自查放行 this file exists against. Exit 0 now means "the gate was bound by the discipline and cleared, and nothing has moved since", never "the review passed".
• No spelling is relaxed. Clause-②: yes / Clause-②: no remain the only two readings; the declaration reader is untouched.
• The tool still writes nothing. No label, ever.

The exit register keeps its shape — 0/4/3(/2), 4 for pair-adverse, never 0-with-a-message — and the docblock's exit table is updated to match. Exit 2 already existed for --pair; it now also covers an unreadable event stream.

Cost

Event streams are fetched only for pairs already in C3's candidate shape (declared yes, bare on both carriers), and the head commit only once both carriers read cleared. needsGateHistory is the single predicate the live reader and the UNJUDGED accounting share, so the set that owes a stream and the set that gets one cannot drift apart. Measured live: full sweep 8s, 9 pairs, 0 UNJUDGED.

Two readings that shaped the implementation:

• The event endpoint returns rows OLDEST FIRST, so a stream read short loses the removal and reads as a live hang or a never-hung gate. The reader pages to exhaustion and hands null — UNJUDGED — when it cannot.
• Head motion is read from the head commit's committer date, one exact uncapped request. The unforgeable alternative (head_ref_force_pushed) exists only on the timeline endpoint, which caps at 250 events and truncates silently; a capped read backing a clean verdict is the fail-open shape this family refuses everywhere else. The residual hole — a force-push landing a deliberately backdated commit — is named at the reading site, and the human PASS conjunct is what closes it.

The gate label constant and the label-event vocabulary are imported from check-half-states.mjs rather than restated, so this row cannot disagree with H31/H35 about what a gate event is. The repo-wide, windowed form of the one-carrier question stays H35's; this row is the per-pair form and is unbounded in time, and it points at H35 rather than re-judging it.

Verification

Reverse-verified in two legs, each mutate → prove on disk → run → restore → prove restored (hash equals the HEAD blob, git diff HEAD empty):

• The card's own measurement, live. Old tool on the real pair — --pair 13864 → exit 4 with the C3 row on card The undeclared-field door sits in FRONT of the hooks, so a key a beforeInsert hook writes has no door at all — and the drivers then disagree (memory stores it, SQL throws a raw statement error) #13657. New tool, same pair, same minute → exit 0.
• Do the new cases bind? Breaking only the completed-state early return turned the self-test red: 2 of 98 case(s) failed, naming the completed-state cases. Restored: 98 pass.

Self-test grew 64 → 98 cases, with fixture pairs for all four states replayed from the 2026-09-01 measurement (card #13657 hung 16:55:54Z / cleared 08:54:47Z; PR #13864 hung 15:19:53Z / cleared 08:54:56Z — nine seconds apart; head 9af92aa3 dated 08:16:59Z). lint.yml still runs the self-test only; the network sweep stays out of CI.

Gates at final HEAD 260a96e4d, exit codes captured before any pipe: the full derived union (dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 15 commands, re-derived at final HEAD and identical) passes, plus repo-wide pnpm lint (eslint . --no-inline-config, exit 0, full scan — no narrowing). check-test-completeness is NOT MEASURED rather than green: it grades a saved turbo run test log that CI tees and no standalone invocation can produce.

scripts/pm/**-only diff — publishes nothing from any package, so skip-changeset rather than a changeset file.

Generated by Claude Code

---

Generated by Claude Code

[os-sam]

Ready + auto-merge armed — provenance per the landing discipline.

• In-seat review verdict: ACCEPT, recorded on the card ([finding] check-clause2-carriers.mjs --pair answers C3 / exit 4 on a LEGITIMATELY cleared clause-② pair — the completed state (declaration outlives the label) is indistinguishable from the fail-open it hunts #14155, comment 5492284866), grounded in a full diff read plus the seat's own re-run of the self-test at head 260a96e4d (98 cases, exit 0 captured directly).
• Landing preconditions verified at arm time: non-governed diff (scripts/pm/** only, not on the GOVERNED_SURFACES register — the Governed Surface Queue Guard check agrees, success); all checks green on 260a96e4d — 32/32 check runs success or skipped, zero failed, zero pending, legacy status success (not a required-subset reading).
• At MERGED: card [finding] check-clause2-carriers.mjs --pair answers C3 / exit 4 on a LEGITIMATELY cleared clause-② pair — the completed state (declaration outlives the label) is indistinguishable from the fail-open it hunts #14155 strips to [priority:p2, finding, domain:skills], assignees cleared.

---

Generated by Claude Code