docs(pm-dispatch): findings pipeline — hold verdicts swap to pm:on-hold, first-touch grading every round, auto-concentrated rounds, lane-evidence protocol, atomic ruling-recording

[hotlong]

Fixes #8449

Implements the findings-pipeline redesign in .claude/skills/pm-dispatch/**, per the card body as amended twice by the maintainer's design input in the card's comments — both amendment comments are in scope and amendment 1 supersedes the original clause 1 (comment-counting metric + exemption comments) with the state swap.

What changed

SKILL.md — 686/686 before and after, strictly net zero (accounting below):

1. State semantics (amendment 1): finding now means exactly "awaiting first grade" — a hold verdict at any grading round swaps finding to pm:on-hold in the same label write; the raw label count is the health metric; the originally-proposed exemption-comment mechanism is dropped and recorded as dropped (do-not-revive guard in both files). The finding and pm:on-hold State-model rows carry the swap and the hold comment's provenance duty (maintainer-ruled vs seat-graded lives in the hold comment, never a second label).
2. First-touch grading every round: budget 3–5 cards, prioritized over re-verifying old holds; the three-way verdict (promote / close-not-planned / hold-swap) and "grading happens at the round, not at filing" stay.
3. Auto-concentrated mode: honest finding count above 15 makes the next fire a domain-batched concentrated round under the 2026-08-13 standing authorization; the sweep bundling five-condition rule applies. The original body's second threshold (non-exempt stock above 30) collapses into the honest count by construction after the state swap and is not restated.
4. Lane-evidence protocol codified: lane seats may attach evidence / premise re-checks to findings in their lane — never grades, never label changes; the skills-lane self-grading carve-out is stated in the findings-round text; the standby-duties line now carries the same no-grade boundary.
5. Atomic ruling-recording (amendment 2): 落卡/升级流程 gains step ③ — recording any ruling (maintainer ruling or delegated adjudication) is one atomic action with four parts: freshness gate / state transition in the same write / Blocked-by: liveness check (a merged half counts as discharge) / condition-fired check. The two mechanical flags are described as report-only half-state-sweeper duties in the runbook; wiring is a separate card and scripts/pm/check-half-states.mjs is untouched (read before deciding: none of its H1–H7 implements either flag today, so describe-here / wire-later is consistent).
6. Round-report metric line: "finding median age" becomes finding count + median age, with the honest-count parenthetical (hold cards have left the label).

references/dispatch-runbook.md (+56 lines; references carry no ratchet) — two new sections host what the hot file sheds: the four verbatim maintainer quotes (dated, untranslated), the state-semantics rationale, concentrated-round batch parameters (15–20 cards, same-family same-context), the sweep five conditions (moved out of SKILL.md verbatim), the lane-evidence boundary, the atomic-recording worked failure shape, the two mechanical flags, the premise re-check command shapes and the escalation analysis template that SKILL.md now points at. The header's source-section list gains 「升级与决策」. Note: the dispatch prompt suggested seat-post-protocol.md as the findings-round host "per its header", but that file's header scopes it to the seat-post protocol / epic teardown / release board; dispatch-runbook.md's header already claims 分诊座位职责, and it already hosts the other triage-seat detail sections — so the detail landed there.

references/seat-post-protocol.md (±2 lines) — the release board's freshness rhythm was keyed to "same cadence as the findings round (every ~5 rounds)"; first-touch grading is now every round, so that anchor went stale and the board self-carries the ~5-round value.

SKILL.md line accounting (ceiling 686, headroom 0)

Region	Before	After	Δ
State model rows (pm:on-hold, finding)	2	2	0 (widened in place)
分类动作 paragraph (finding wording, tighter wrap)	7	6	−1
发现分诊轮 paragraph (full rewrite; five-condition detail moved to runbook)	6	6	0
发版板 freshness anchor	2	2	0
Round-report health metrics	2	2	0
Standby findings-duty line	1	1	0
落卡/升级流程 (①② rewrapped, analysis template moved to runbook, new atomic step ③)	6	7	+1
Total	686	686	0

The +1 (atomic four-parter) is funded by the −1 (分类动作 rewrap); the findings-round rewrite funds itself by moving the sweep five conditions and the ruling quotes to the runbook.

One-time migration — enumerated only, ⛔ not executed here (amendment 1)

Amendment 1 sized the migration at ~26 held findings out of ~47. Measured at implementation time (2026-08-14T01Z), the number drifted hard, in two ways:

1. The bulk of the migration was already executed by another actor. 61 cards carry pm:on-hold today, and a tight cluster of formerly-finding cards shows updated_at 2026-08-13T13:49–13:50Z — minutes after amendment 1 was posted, hours before this card was dispatched. The amendment's sequencing ("the migration rides the skill PR's landing") was overtaken on the live board. Per the shared-identity rule I did not revert or re-label anything; this is reported as an observation.
2. 9 open finding cards remain. By latest triage comment:
   • The #8497 MCP write-response tripwire enumerates the stdio bridge only — the runtime's HTTP McpDataBridge is a second implementation no guard walks #8589 — HELD (graded 2026-08-14, three named re-grade triggers) → migrate
   • finding: check:engine-double-contract counts declaration sites, so a behaviourally-distinct engine double built by Object.assign over an existing one is not counted #8553 — HELD (graded 2026-08-13, three named triggers) → migrate
   • [finding] dispatch-gates hintCovers matches on a raw STRING prefix, not a path-segment boundary — a hint naming one package can claim a sibling that merely shares its name prefix #8534 — HELD (graded 2026-08-13 by the skills seat, trigger-fired re-verification stated) → migrate
   • plugin-sharing re-derives materializeDeclaredFields instead of importing it — a second copy of the declared-field binding contract, already diverged #8489 — HELD pending a PR that has since MERGED (2026-08-13T18:14Z): the named restart already fired → re-grade queue, not migration (exactly step ③'s condition-fired check)
   • A licensed max_nodes cap still cannot be enforced by any replica: the cluster has no membership view and no slot claim, so a count-carrying gate verdict stays advisory #8501 — held pending a blocker discharged the same day; latest comment states it is still ungraded → re-grade queue
   • [finding] GitHub MCP list_issue_types returns 403 for this integration while issue_write sets native types fine — a seat that probes the list first will wrongly conclude typing is unavailable #8574 — PROMOTED riding another card's dispatch; keeps finding as the sweep-member record
   • finding: the declared-index replacement arm's guards are redundantly covered, so any single one can be deleted with every test still green #8557 — routing comment explicitly defers grading (lane evidence attached, no grade) → first-touch queue
   • finding: the #8435 detector cannot reach a correctly-refusing gate whose testimony is governance — two independent stops, measured on check-driver-memory-census.mjs #8585, publish-drafts has two more undeclared-driver-text fields on the same 200 body: unhideError and rebindError #8516 — zero comments → first-touch queue

Migration to execute after this merges: #8589, #8553, #8534 — one label write each, finding → pm:on-hold; hold comments are already on-card. Post-migration honest finding count: 6.

Verification

• pnpm check:pm-skill-ratchet — green: "SKILL.md is 686 lines (ceiling 686; headroom 0)"
• pnpm check:pm-skill-id-lint — green: 9 files clean (no issue-ID citations in operational text)
• pnpm check:skill-frame-sync — green: 4 copies isomorphic, 3 axes, 4 count mentions agree (edits stay clear of every anchor)
• pnpm check:skill-frame-freshness — green (frame unchanged vs origin/main)
• pnpm check:nul-bytes — green (7740 files)
• pnpm check:doc-authoring — green (377 files)
• pnpm --filter @objectstack/lint run check:doc-formula-expressions — green after building the lint dependency closure (first run was the documented fresh-worktree false red: unbuilt @objectstack/formula dist)
• Gate re-derivation: node scripts/pm/dispatch-gates.mjs over the actual changed paths returns exactly the five families the dispatch prompt named (plus the any-edit check:nul-bytes); no family surfaced beyond the prompt's list; check:skill-frame-freshness run additionally as the implicated sibling of the named frame-sync family.

Constraints honored

• Model tiering section, the queue/dispatch sections landed by the strategy PR, and scripts/pm/check-half-states.mjs: untouched.
• Maintainer quotes recorded with date + verbatim original, untranslated, in the runbook (SKILL.md cites by date and points there — the same pattern the 选择优先级 section already uses).
• ADR-class merge path (skills root): left as draft for the maintainer's hand-merge — not marked ready, no auto-merge, no queue.

One seam observed and left open (also in the dev report): the sweep five-condition rule keeps finding on promoted sweep members (#8574 is a live instance), which re-pollutes the honest ungraded count while a sweep is in flight. Not resolved here — the five conditions are a 2026-08-07 maintainer approval this card does not reopen.

---

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 1:04am

[hotlong]

同意合并

[hotlong]

Audit — landing on explicit per-PR maintainer authorization. The maintainer's review on this PR (2026-08-14 04:14Z, verbatim, untranslated): 「同意合并」. That satisfies the human-review requirement for skills-root changes (founding ruling 2026-08-11: 「所有 skills 的更新和 adr 类似,需要人工审核」 — the review is the human act; this seat's ACCEPT of record is on the card). Acting on it via the only sanctioned landing path: ready-flip + auto-merge (SQUASH) into the merge queue — ⛔ no direct merge. Executed by the skills seat, session session_018WuTtyckQa1VcXwgd52JpN; on MERGED, the card's close-out chain runs (three-card migration + serialized dispatches).

---

Generated by Claude Code