pm-dispatch: the findings pipeline is structurally under-drained — switch the metric to UNGRADED count, add first-touch grading, auto-concentrated rounds, and codify the lane-evidence protocol

[hotlong]

Filed by the triage seat Routine on the maintainer's direct instruction (2026-08-13, live session on the triage fire). Maintainer's words, verbatim:

「Finding 很多,是否建议集中分诊」 … 「按照你的建议」 … 「还有这个积压时常态,是否应该优化。」

The first two authorize domain-batched concentrated finding rounds (executing from 2026-08-13); the third asks for the structural fix — this card is that fix, scoped to .claude/skills/pm-dispatch/**.

Measured state (2026-08-13)

• 47 open finding cards vs the ~15 insertion trigger — the stock has been ~3× the trigger for days, i.e. the trigger fires every round and has stopped carrying information.
• Inflow ~10/day (dev seats' out_of_scope_findings, by design — 发现快于关闭是循环在工作; do NOT suppress inflow).
• Historical outflow: 0–5 grades/round, discretionary; several rounds did 0–1.
• The stock is heterogeneous and the raw count conflates three very different populations: ungraded (~15 today — never had a first verdict; premise accuracy decays in days), actively-held (graded, restart conditions live, needs periodic re-verification), and stable-ruled holds (maintainer-ruled or verdict-stable; e.g. the card ruled deliberate-defer on 08-11 was explicitly exempted from per-round re-verification on 08-12 because four consecutive passes returned identical results against an unchanged file).

Root causes

1. The metric counts the wrong thing. Total open findings includes stable holds that are correctly parked; a healthy holding pen reads as a crisis, and a real ungraded pile hides inside it.
2. No first-touch SLA. A new finding is only read when a findings round happens to pick it up; time-to-first-grade is unbounded, and stale premises are the single biggest per-card cost (re-verification archaeology).
3. Re-verification of stable holds eats the budget that ungraded cards need — the 08-12 exemption precedent fixed this for one card; it is a pattern, not a one-off.
4. Grading is a single-seat bottleneck by design (single-producer is correct), but lane seats already produce the expensive half (evidence, premise re-checks) informally — the drivers-seat evidence comment that completed a promotion this week is the model. Uncodified, it happens by luck.

Proposed skill changes (4 clauses, for the dev to land in the skill text)

1. Metric swap: the round brief's finding-health line reports ungraded count (open finding with no triage-verdict comment) and median age of ungraded; total stock becomes a parenthetical. Stable-ruled holds get an explicit exemption comment naming re-open triggers (the 08-12 pattern) and leave the per-round polling surface.
2. First-touch grading: every round (not every ~5), the sweep grades findings with zero triage comments, budget 3–5, prioritized over re-verifying old holds. Re-verification of active holds keeps its existing cadence.
3. Auto-concentrated mode: ungraded >15, or non-exempt stock >30 ⇒ the next fire runs a domain-batched concentrated round (15–20 cards, grouped so same-family cards are read in one context; the sweep-打包晋级 five-condition rule applies). Standing authorization: maintainer 2026-08-13 (quoted above) — no per-instance ask needed.
4. Lane-evidence protocol, codified: lane seats MAY attach evidence/premise re-checks to findings in their lane (⛔ no grade, no label change — grading stays triage's); the skills-lane carve-out (that seat grades its own lane's findings) is unchanged and should be stated where the findings-round text lives, so repo-wide triage doesn't double-produce there.

Constraints for the implementing dev

• ⛔ scripts/pm/check-skill-line-ratchet.mjs: SKILL.md only shrinks. New clauses need compensating trims in the same area (the existing 发现分诊轮 paragraph is the natural host — this is a rewrite of its trigger/budget sentences, not an addition), or a maintainer ratchet-raise quote in the PR body.
• ⛔ No issue-number references in operational text (check:pm-skill-id-lint) — the exemption/evidence patterns must be described self-contained, not cited.
• Mandatory model clause applies: this card edits .claude/skills/pm-dispatch/** ⇒ dispatch with claude-fable-5, no downward discretion. Resulting PR is human-merge class (skills root): reviewed on-issue, left visibly unmerged, listed as awaiting a human merge.

Dedup: searched open cards for finding-backlog/threshold/concentrated-round shapes — the shift-end/hot-handover card and the dispatch-strategy-four-items card are adjacent but neither touches the findings pipeline; no overlap.

[hotlong]

Scope amendment from the maintainer's design review (2026-08-13, live session) — this supersedes clause 1's metric mechanism. Maintainer's words, verbatim:

「积压数字本身有误导 —— 47 张里大半是带重启条件的健康 hold(它们本来就该躺着),那为什么标 finding」

The question exposes the real defect: finding currently encodes TWO states — "awaiting first grade" and "graded, verdict = hold with restart conditions" — which violates the label-is-the-state-machine honesty rule. Amended design, replacing the original clause 1 (comment-counting metric + exemption-comment mechanism):

1. A hold verdict at any grading round swaps finding → pm:on-hold (domain label kept; the hold comment carries date / reason / named restart conditions, as pm:on-hold already requires). A seat-graded hold is a decision made within the seat's discretion, which is exactly pm:on-hold's declared semantics; maintainer-ruled vs seat-graded provenance lives in the hold comment's citation, not in a second label.
2. finding thereby means exactly one thing: observation-class, awaiting first grade. The finding-health metric becomes the raw label count — no comment-counting predicate needed. The exemption-comment mechanism this card originally proposed is DROPPED: state transition replaces it (a stable hold leaves the finding-round polling surface by construction).
3. Graded holds inherit pm:on-hold's existing reader paths (dispatch-time trigger-file checks; restart conditions that name concrete triggers). Periodic premise re-verification stops being a per-round tax on every hold and happens when a named restart fires — which is the cost model the maintainer's question implies.
4. One-time migration rides the skill PR's landing, not this card's filing: the ~26 currently-held findings (grade comment present, verdict = hold) swap finding → pm:on-hold, one label write each, hold comment already on-card. After migration the finding count is the ~10 skills-lane ungraded cards — honest on day one. The migration list is derivable mechanically (open finding cards whose latest triage comment says HELD); the implementing dev should enumerate it in the PR body for the record, ⛔ not execute it before the protocol text merges.

Clauses 2 (first-touch grading), 3 (auto-concentrated mode) and 4 (lane-evidence protocol) of the original body stand unchanged, except thresholds now read against the honest finding count.

---

Generated by Claude Code