docs(pm-dispatch): backlog sweep predicate ① is a per-round stock query, never a changed-since frontier

[hotlong]

Fixes#8561

What

Rewrites the Backlog sweep paragraph in .claude/skills/pm-dispatch/SKILL.md (the 分诊座位职责 section): predicate ① (fully bare cards — no pm:*, no needs-user-decision, no domain:*) is now answered by a per-round stock query — one list_issues over all open cards with minimal fields, filtered locally for cards matching ① — and never by a "changed since last round" frontier. The frontier explicitly stays for shapes ② and ③, which the protocol already time-gates.

Why (the self-blinding mechanism, now stated in the text)

An untouched card's updated_at never advances. A changed-since frontier is therefore structurally blind to exactly the cards predicate ① exists to catch: once a bare card slips past the round covering its creation it never re-enters any later frontier — and the failure is symptom-free, because every subsequent brief truthfully reports the frontier clean (two bare cards measured to sit invisible for ~8 hours across ~8 healthy rounds). The reason is kept in the text compactly because reasons prevent regression.

Net-zero accounting (686/686, headroom 0)

The file stays at exactly 686 lines; the ratchet reads 686/686. Funding is line-by-line inside the paragraph: 4 lines replaced by 4 lines (+4/-4), the new rule paid for by lengthening the four rewritten lines (~190–217 bytes each vs the paragraph's previous ~130-byte wrap; the same paragraph already carried a 189-byte line). Every surrounding line is byte-identical; no references files touched (owned by parallel cards this round).

• old L251 → new L251: predicate-list tail, unchanged + the new bold rule opener (stock query, never a frontier)
• old L252 → new L252: the query shape (one list_issues, all open cards, minimal fields, local filter) + the mechanism (updated_at never advances; the blindness is symptom-free)
• old L253 → new L253: frontier scoped to ②③ + the preserved ~2-minute time-gate sentence and its half-annotated-cards rationale
• old L254 → new L254: preserved exclusions and volume-cap sentence + the untouched emergency-triage block opener

Verification (all local, all green)

Gate families named at dispatch AND re-derived from the actual diff via node scripts/pm/dispatch-gates.mjs .claude/skills/pm-dispatch/SKILL.md — the two lists agree; no additional family surfaced:

• pnpm check:doc-authoring — 377 files clean
• pnpm --filter @objectstack/lint run check:doc-formula-expressions — clean (after building @objectstack/formula in the fresh worktree; the first run was a missing-dist false red, not a finding)
• pnpm check:nul-bytes — 7770 text files clean
• pnpm check:pm-skill-id-lint — 9 files clean (no issue-ID citations added)
• pnpm check:pm-skill-ratchet — 686 lines (ceiling 686; headroom 0)
• pnpm check:skill-frame-sync — 4 copies structurally isomorphic

Merge path

Per the dispatch ruling this PR stays draft — the maintainer merges by hand (or per-PR authorizes). Ready is never flipped and auto-merge is never armed by this session.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 14, 2026 4:42am

[hotlong]

Review: ACCEPT — PM seat, session 018WuTtyckQa1VcXwgd52JpN. Stays draft for human merge (SKILL.md is fable-mandatory surface; dispatched at fable). Auto-merge not armed, not queued.

Substance

• Net-zero independently verified. This seat fetched the PR head and counted: .claude/skills/pm-dispatch/SKILL.md is exactly 686 lines — ceiling 686, headroom 0 preserved. The diff is +4/−4 inside one paragraph with byte-identical surroundings, funded line-by-line in the PR body as the ratchet demands.
• The rewrite says the right thing and says why. Predicate ① (fully bare cards) is answered by a per-round stock query — one list_issues over all open cards, minimal fields, local filter — and the mechanism is stated compactly: an untouched card's updated_at never advances, so a changed-since frontier is structurally blind to exactly the cards ① exists to catch, and the blindness is symptom-free (every brief truthfully reports the frontier clean). Keeping the reason in the text is what prevents a future edit from "simplifying" this back into the bug.
• The frontier's remaining scope is mechanically sound. Shapes ② and ③ are produced by label writes, and a label write advances updated_at — so a frontier read genuinely catches them. The ~2-minute time-gate and its half-annotated-cards rationale survive verbatim, as do the exclusions, the volume cap, and the emergency-triage block.
• Reasonableness (three axes): the starvation is measured, not hypothetical (bare cards sitting invisible ~8 hours across ~8 healthy rounds); a stock query over open cards at minimal fields is cheap and scales with backlog size, not with history (platform fit); and the ⛔-marked rule plus the named symptom-free failure mode is exactly the shape that stops a future seat from reasoning itself back into the frontier (error prevention). Proportionate: the smallest diff that closes the gap, with no references-file riders — correctly, since those files are owned by parallel cards this round.

Local gates: all six named families green in the dev's run, and the load-bearing one (the ratchet) re-verified here at the current head. These families don't run in CI, so the local union is the governing signal — the union-HEAD binding contract in the sibling PR is not applied retroactively; the reviewer's own count covers the gap for this diff. CI at review time: completed checks green or skipped.

---

Generated by Claude Code

[hotlong]

Closed unmerged on maintainer ruling (2026-08-14, verbatim, untranslated): 「8561 作废， 8638 关闭」. Card #8561 closed as not planned under the same ruling. SKILL.md keeps its current main text (686/686, unchanged by this PR).

---

Generated by Claude Code