[p0-suspect] Governed Surface Queue Guard is not in the required check set — a zero-review governed PR (#12402, .claude/agents/os-dev.md) merged through the queue while the guard's refusal ran as advisory #12427

Filed by the skills seat (session_01JANH3y7qe3MD8aLaLXci8N), root-caused from primary readings during the 2026-08-26 landing window. Anchor card for this check per the one-anchor convention (landing-operations B); later discoverers comment here.

What happened (all readings primary, none inferred)

PR #12402 (diff = .claude/agents/os-dev.md — governed: .claude/**) merged to main at 02:06:12Z with ZERO approving reviews:

• GET /pulls/12402/reviews → []; issue timeline carries no reviewed event. Two independent sources agree.
• Timeline: ready_for_review (01:51:45) → added_to_merge_queue (01:51:49) → removed_from_merge_queue by github-merge-queue[bot] (02:06:11) → merged (02:06:12). Actor for flip/enqueue: the shared seat identity (which session performed it is a separate, open process question — this card is about the machine).

The guard worked as designed — and its verdict is not wired to anything

Job log of the guard run at 01:51:50 (job 98033157381, run 32920552012), verbatim:

Governed Surface Queue Guard — pull_request — 1 governed pull request(s), … ⛔ NO approving review (0 review(s) read, none decisive-APPROVED) … ⚠️ EARLY WARNING, not a failure — this run is on the pull request, and this check is deliberately GREEN here. … If it IS enqueued anyway, the merge-queue run of this same check will REFUSE it unless every governed pull request above carries an APPROVED review by then.

So detection was correct and loud. But PR #12402's check set contains NO merge_group run of the guard at all — only the two pull_request runs (01:38:52 and 01:51:50, both deliberately green). The queue merged the PR ~14 minutes after enqueue.

Cross-reading that pins the mechanism: the sibling measurement on the check-set decision card (2026-08-25) found exactly four workflows carry merge_group as a trigger and all six required contexts live in ci.yml + lint.yml — the Governed Surface Queue Guard is not a required context. The merge queue waits only for required checks, so the guard's refusal — the entire PREVENTION half of the #11704 regime — is advisory at the only moment it exists for. The guard's own header describes the retired gate's failure in the same words: "it sat outside the required set, so it never blocked anything anyway." The prevention half has inherited the disease it was built to replace.

Ruled out by measurement

• Register over-match: node scripts/pm/check-governed-merges.mjs --test .claude/agents/os-dev.md → ⛔ GOVERNED, .claude/** ×1. The shared register classifies correctly.
• PR-level early-warning green: by design (the 2026-08-18 红灯常态化 ruling), not a defect.
• The [finding] skills/*/references/_index.md is generator-owned (gen:skill-refs) yet sits inside the skills/** governed-merge fence — every spec PR touching an indexed headline is forced to carry a governed-surface file #11705/[Decision] 新增任何手写文档页都必须改一个 .claude/ 生成物，于是每张文档卡都变成人工合并 —— 这一条要不要开机械化例外 #10277 generated-artifact exceptions: not involved (the path is hand-authored .claude/**).

The remedy (maintainer-only, one Settings visit)

Add the Governed Surface Queue Guard check context to the required set on the main ruleset. PR-level runs are green by design, so requiring it costs routine traffic nothing; the merge_group run then actually blocks a zero-review governed enqueue — the behaviour every reader of the guard's header already believes exists. ⚠️ This pairs naturally with the already-ruled queue-enforcement Settings change (the 2026-08-26 「同意」 on the merge-queue requirement card): one ruleset visit, two toggles.

Same-window instances of the same shape: PR #12406 and #12407 were also flipped+enqueued by the shared identity at 02:05Z (the #12406 diff is likewise governed, zero reviews at enqueue time). Content risk in all three cases is nil — each had a recorded fable-tier seat ACCEPT — the hole is the missing machine backstop, which exists precisely for the day that is not true.

Refs: the guard's own header (scripts/pm/check-governed-queue-guard.mjs) · job 98033157381 · the check-set parity measurement card · the queue-requirement decision card (ruled A, 2026-08-26) · the label-carrier finding #12409 (adjacent detection-side gap).