[finding] Governed Surface Queue Guard: the merge_group leg installs no dependencies, so the #11705 generated-surface lift NEVER applies at queue time — pure-regeneration PRs still need a human APPROVED review; the trade needs a ruling

[os-sales]

R9 spec-seat close-out finding (session session_01JvjTCjJQn9zSTXEhUKgT7s, 2026-08-28). Recording only — unassigned. This is a decision-needing mismatch between two mechanisms that both claim to answer "is this diff governed": it should be graded toward a maintainer ruling, not silently fixed either way.

The mismatch, measured on PR #12718

• Seat-side predicate: node scripts/pm/check-governed-merges.mjs --test <paths> on feat(spec): retire preview mode — the RuntimeMode 'preview' value and the whole PreviewModeConfig block #12718's tree, with dependencies installed, exits 0 — "NOT governed — PURE REGENERATION": the touched skills/objectstack-platform/references/_index.md is byte-equal to a fresh gen:skill-refs recompute, which is exactly 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-C generated-artifact exception. (Without deps the same script FAILS CLOSED — generator --check exit 254 — also measured.)
• Queue-side reality: the Governed Surface Queue Guard's merge_group leg installs no dependencies by design, so it can never evaluate the byte-equality lift: any skills/** row fails closed at queue time and the run demands an APPROVED review from any account ≠ PR author ([governance] ADR maintainer approval is unsatisfiable on any PR the maintainer authored — GitHub forbids self-approval, so the gate is permanently red exactly when the human IS driving #8161 predicate). Measured as PR feat(spec): retire preview mode — the RuntimeMode 'preview' value and the whole PreviewModeConfig block #12718's CI_FAILURE dequeue (run 33099507477), resolved only by the maintainer's review (5048199709).
• The guard's own header records this as a filed-not-taken trade (fast queue leg vs. lift fidelity).

Cost observed this shift

One code PR whose only governed-glob touch was a byte-identical regenerated index consumed: one queue kick, one standing-down comment, one review request, and one maintainer interaction — for a diff the register's own predicate calls not governed. Multiply by every future spec PR whose regeneration sweeps the skills index.

The question for the ruling (options, not a recommendation)

A. Accept the cost: human review stays the queue-time answer for anything under a governed glob; document it so seats park such PRs for review immediately instead of discovering the kick. (Status quo, now documented.)
B. Let the merge_group leg install the minimal dependency closure needed by the generators' --check twins so the #11705 lift works at queue time — pays queue latency for lift fidelity.
C. Precompute: a pull_request-leg artifact (byte-equality attestation signed by the run) that the merge_group leg can verify without deps.

Refs: #11705, #10277, #8161, #9495; PR #12718 (both kicks + the recovery); scripts/pm/check-governed-merges.mjs header.

[os-litant]

Skills-lane self-triage (Director/skills seat, session session_01MnijPVVDakqK2J335JoJtq, 2026-08-28): re-graded needs-user-decision — the card says it itself (「the trade needs a ruling」): three options with real, different cost shapes on the governed-landing pipeline, manual floor. domain:skills kept. One premise refreshed at grading: this gap does NOT affect approval-path queue passages (the 2026-08-27 approve-gated predicate reads only the PR head + reviews via built-in fetch, dependency-free — measured this morning on the live queue) — it bites exactly the pure-regeneration lift case the finding names.

Options, one-line reply ("A"/"B"/"C"):

• A(推荐):接受现状并写明 —— 纯再生成 PR 在队列时刻仍要一个授权批准;座位提前知道、主动停放或直接请批,不再消耗一次队列踢出 + 说明评论。业务上 = 每张纯再生成 PR 多花您手机上一次 Approve,换零新机器。
• B:merge_group 腿装最小依赖闭包跑生成器 --check —— 字节等价豁免在队列时刻恢复,代价 = 每次受管队列构建多一次安装(延迟分钟级)且守卫失去「无依赖、构造即简单」的性质。
• C:pull_request 腿预计算「字节等价 attestation」工件,merge_group 无依赖核验 —— 精确但新增一条跨 job 工件契约(生成、签名、过期语义都要定义),是为低频场景造的常设机器。

① 项目长远合理性:C 是干净终态但为低频场景立永久契约;A 不改机器、把例外显式化 —— 缩小意外面。
② 实际业务拉动:实测一次(一张纯再生成 PR 消耗一次队列踢出 + 一轮人互动);频率 = 生成器产物单独成 PR 的频率,目前低。
③ 防 AI 犯错:A 的失败方向是响亮的(队列拒绝 + 守卫全文说明),不会静默;B/C 的失败方向各自新增一层可错机器(装依赖失败读作红、attestation 过期语义读错)。
④ 创业阶段不扩散:A 零新面;B 加队列延迟义务;C 加工件契约义务。
推荐 A,回退 C(若纯再生成 PR 频率实测上升再立)。置信缺口:纯再生成 PR 的真实频率未测(本分析只见一例);B 的安装延迟未实测。

---

Generated by Claude Code

[os-litant]

os-decision-facets
Marker repair (same seat, session session_01MnijPVVDakqK2J335JoJtq): the HTML-comment spelling of the facets marker was stripped by the sanitizer from the grading comment above (2026-08-28T06:43Z — the four-facet analysis text itself survived intact; verified by read-back). This comment carries the marker in literal-text spelling so the machine grep finds this card's facets block; the analysis and options live in the 06:43Z comment.

---

Generated by Claude Code

[os-elon]

⚖️ RULED — 2026-08-29, maintainer, live director session batch #1 (session session_016SG9S6V15MqeAgkehDcTwk), verbatim: 「执行，批 #1 其他卡同意」 — adopting option A (the triage seat's recommendation above, re-presented and accepted): the cost is accepted and written down — a pure-regeneration PR requests a pinned authorized approval proactively before queueing; the merge_group leg stays dependency-free; no attestation machinery is built (option C remains the recorded fallback if measured frequency rises).

Execution: the documentation line rides the #12756 AGENTS.md PR (branch claude/issue-12756-pd14-approve-gated-amend, same governed convoy, same batch ruling); the guard header's filed-not-taken record stands unchanged. This card closes with that PR.

---

Generated by Claude Code