docs(pm-dispatch): clause-② negative boundary + attach-time rule for needs:contract-review

[os-elon]

Fixes#12887

Governed protocol change — this PR touches .claude/** and stays DRAFT for the maintainer's hand-merge (Prime Directive 14, 「人工合并即人工审核」). Never ready, never queued, never auto-merge, by any seat.

Executes points 1, 2 and 4 of the maintainer's 2026-08-28 ruling (live director session, recorded on the card; adoption verbatim and untranslated):

同意，你现在就执行，并派发这个卡片

Points 3 (no new label, no rename) and 5 (stock cleanup) were already executed by the director seat before this dispatch: all 18 pre-marked open carriers stripped with per-card provenance comments — migration state at dispatch time: label:needs:contract-review is:open = 0. An open carrier of the label is from now on always a real pending review.

What the text now says

1. Negative boundary on the clause-② criterion (SKILL.md, 条款② clause): a runtime permission/security behaviour change is NOT clause-② — it is the 安全/权限边界 manual-floor category, whose control is the ruling itself, not the review chain; clause-② is the published contract face only. Evidence note carried in place: two independent seats derived the same wrong reading from the old text (a third seat self-audited the same conflation on the card).
2. Attach-time rule (SKILL.md, 条款②入队闸门 paragraph): needs:contract-review attaches only when a reviewable contract increment EXISTS (draft PR, or a report arriving first); forward pre-marking on queue/blocked cards is retired. The forward clause-② fact lives in the three existing mechanisms — the card's ruling/triage comments, the claim's mandatory Clause-②: yes|no declaration, and dispatch-gates --tier output.
3. The ⛔ 不预挂 line (references/contract-review.md, 载体纪律): the same rule at the operational layer, superseding the triage "pre-hang on queue cards" practice; it points at the main file's gate paragraph for the three channels (the reference's own header already declares the main file as the principles home, so the text stays self-contained).

Zero-headroom ratchet: net 0 per file, cut ledger

Ratchet verdict lines (pnpm check:pm-skill-ratchet, run at 04c7536, exit 0):

✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md: widest table row is 765 bytes (pin 765; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 1005 lines (ceiling 1005; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/contract-review.md: widest table row is 0 bytes (pin 0; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/contract-review.md is 48 lines (ceiling 48; headroom 0).

Mechanical proof over the whole payment: for both files, render-normalized OLD text (CSS segment-break transformation: a break between two East Asian wide characters renders as nothing, any other break as one space) plus the two ruled insertions and the five cuts below equals render-normalized NEW text byte-for-byte. So every other visible diff line is pure re-flow of sloppy wraps (orphan lines of 3–46 bytes), moving whitespace only. The literal spaces that appear inside re-flowed lines (e.g. before 才落) are pre-existing renders made visible — the self-concealing re-wrap mechanism documented in the ratchet script's header — not introduced spaces. New lines were wrapped with the ratchet script's own exported wrapLine (the canonical wrapper the gate verifies).

SKILL.md (adds: negative boundary +266 bytes, attach-time rule +292 bytes; paid entirely by re-flow, net 0 at 1005/1005):

• Re-flow, model-tier paragraph (6 → 4 lines): three orphan wraps of 10/10/22 bytes merged back into their sentences. No content cut.
• Re-flow, clause-② paragraph (16 → 17 lines including the +266-byte insertion): absorbs the 46/10-byte orphan wraps. No content cut.
• Re-flow, 条款②入队闸门 bullet (11 → 13 lines including the +292-byte insertion): absorbs the 37-byte orphan wrap. No content cut.
• Re-flow, os-regen bullet (4 → 3 lines): the 9/6-byte orphan wraps merged. No content cut.

references/contract-review.md (adds: the ⛔ 不预挂 bullet, +2 lines / 225 bytes; paid by five same-file cuts, net 0 at 48/48):

• Cut 「;挂与清皆按标签纪律的 read-modify-write 硬步骤写」 (载体不迁移 bullet; frees 1 line) — surviving home: SKILL.md 标签纪律 bullet, which states 「承载闸门语义的标签(如 needs:contract-review)挂与清两向同此四步」; the reference's clause was a duplicate pointer.
• Cut 「曾构造上不可测量,今」 (转录档位核验 bullet) — historical context; surviving home: git history (this file's previous revision) and the 2026-08-27 ruling card. This is the one cut class that touches text near a ruling attribution (the attribution itself — date and quote — is untouched); flagged for your eye.
• Cut 「(重做或维护者路径)」 — post-invalidation remedies; surviving home: the same bullet still mandates invalidation, and the FAIL/REWORK dispositions live in 清标即落地 (「FAIL / REWORK 原样」).
• Cut 「,采信 = 证据清单核对,非实质重判」 — surviving home: the same sentence's 「逐字采纳或整体作废(理由仅机械三种…)」 already states that adoption is mechanical.
• Trim 「父会话对裁决仅两个合法动作」 → 「父会话仅两个合法动作」 — the object is the sentence's own subject; no rule content moved. (The three micro-cuts together free 1 line.)

Gates (run in the worktree at 04c7536; every exit captured before any pipe)

• pnpm check:pm-skill-ratchet — exit 0 (self-test pass); verdict lines quoted above.
• pnpm check:pm-skill-id-lint — exit 0: "✓ check-skill-id-lint: 23 file(s) clean (pattern /#[0-9]{3,}/g)." — the new lines carry the ruling date, no card ids.
• pnpm check:skill-frame-sync — exit 0: "✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files" (frame untouched).
• pnpm check:skill-frame-freshness — exit 0: "✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now)."
• pnpm check:pm-governed-prose — exit 0: "✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces (docs/adr/** · .claude/** · skills/** · AGENTS.md · CLAUDE.md) and claim no others."
• pnpm check:pm-governed-merges — exit 0 (self-test, 206 assertions).
• pnpm check:doc-authoring — exit 0: "✓ doc authoring guard: 392 files clean — no bare metadata literals."
• pnpm check:agent-test-spelling — exit 0: "✓ sweep is clean".
• pnpm --filter @objectstack/lint run check:doc-formula-expressions — exit 0: "✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 425 files / 1453 TS blocks judged clean" — after building @objectstack/formula and @objectstack/lint (the first run refused with PREREQUISITE NOT MET, which per the gate's own text is NOT MEASURED, not a finding).
• node scripts/pm/check-governed-queue-guard.mjs --self-test — exit 0 (113 cases). The full guard is a CI-context gate: it reads the workflow event payload, and a bare local run correctly refuses ("could not read GITHUB_EVENT_PATH" — could-not-look must never exit 0 there). The PR's own CI run is the real judgment for that family.
• pnpm check:nul-bytes — exit 0: "check-nul-bytes: OK (scanned 7168 text file(s) … no raw ASCII control bytes)."

Gate union derived by node scripts/pm/dispatch-gates.mjs with no paths passed (change set derived from git: 2 paths vs merge base 93a554d): 9 matched families — all run — plus the dispatch-named frame-freshness and nul-bytes.

No changeset: internal instruction text under .claude/skills/**, nothing publishes — skip-changeset label applied (this repo's real mechanism). content/docs/releases/ untouched. The previously pre-marked cards named in the card's triage self-audit are the director/triage seats' to re-walk under the now-written rule; out of scope here: #12597 (it remains exactly as the director seat left it).

Generated by Claude Code

---

Generated by Claude Code

[os-elon]

Auto-merge armed by the director seat (session session_016SG9S6V15MqeAgkehDcTwk) — provenance for the governed-merges audit.

The PR body above pledges draft-only + hand-merge; that text was written before the maintainer acted. What changed, in order:

1. 07:52:42Z — os-zhuang approved, review pinned to exactly head 04c7536 (the only commit; no push since).
2. 07:52:45Z — the PR was flipped ready (not by this seat).
3. This comment's timestamp — this seat armed auto-merge.

Basis: the approve-gated queue path for governed surfaces (merged as 93a554d7, this PR's own merge base): a governed PR passes the Governed Surface Queue Guard when a maintainer approval is pinned to the exact queued head. The approval is the human review record; arming auto-merge is the mechanical follow-through, not a seat merging a governed surface on its own authority. If the head moves for any reason, the approval un-pins and the guard fails closed — re-approval required.

---

Generated by Claude Code