pm-dispatch: finding triage runs to empty each fire, quality held by ≤5-card same-family batches

[os-zhuang]

Closes#11972

Governed surface (.claude/**) ⇒ draft, human merge. Verified at a879b1e57.

The ruling

Scheduled finding triage no longer carries a per-round total budget. Every fire
grades all ungraded finding cards; the rubber-stamp risk is controlled by
batch granularity instead of by capping the round. Maintainer, 2026-08-25,
verbatim and untranslated:

单轮跑数小时没关系啊。不读卡的批量盖章 可以通过每批5张这种约定来解决

So: ≤5 same-family cards per batch, read one by one, and a run of several
hours is an acceptable shape. A fire that exhausts its window resumes from where
it stopped — graded cards are off-label and never re-read, so the label
discipline already is the checkpoint.

What changed

.claude/skills/pm-dispatch/SKILL.md, 「发现分诊轮」

• 首触定级每轮跑:预算 3–5 张 → 每 fire 跑到清空: no round total, all ungraded
  findings graded this round, first-touch still ahead of old-card re-verification.
• Quality clause added: 质量靠每批 ≤5 张同族卡逐张读,不靠封顶总量, with the
  2026-08-25 ruling quoted verbatim at the point where the rule is stated.
• The finding >15 ⇒ 集中轮 trigger sentence is gone; 域分批 and the sweep
  promotion five rules stay, now as standing practice rather than a mode switch.
• Unchanged: 零定级评论优先, 判级在分诊不在立单, the skills-lane self-triage
  exception, the Restart-when: daily sub-round, pm:retriage per-fire priority.

.claude/skills/pm-dispatch/references/dispatch-runbook.md, 发现分诊轮细则

• 自动集中轮批量参数:15–20 张 → 每批约定:≤5 张同族卡、同上下文逐张读
  (维护者 2026-08-25,取代 15–20/批). Standing authorization unchanged.
• New 积压告警 bullet, the >15 trigger's destination: 连续 COMPLETED 清空 fire 后仍 >15 ⇒ 入流 > 吞吐,报维护者。 With run-to-empty as the steady state a raw
  count is no longer a mode switch — it is only an alarm once it survives
  consecutive completed fires, which is precisely the test that separates inflow
  above triage capacity from a spike the next fire clears.
• 旧 hold 重验不占预算 removed — 「预算」 no longer exists; the ordering it
  encoded now reads 首触定级…优先于旧卡重验 in SKILL.md.
• Section heading provenance now reads 2026-08-13 / 08-25.

Why run-to-empty rather than a larger budget

The current pile-up is an outage accumulation, not steady-state inflow.
Maintainer, 2026-08-25, verbatim and untranslated:

现在挤压这么多,是因为有一段时间我没token停了分诊,并且改成了两小时一次

That is the strongest argument for the ruled shape: a per-round budget never
recovers from an outage — it amortizes the backlog over dozens of fires while
new cards keep aging — whereas run-to-empty self-heals on the first fire after
supply returns. It is also why the alarm is calibrated on survival across
completed fires: an outage gap is expected to produce a spike and one long
recovery round, and that is by design, not an alarm.

This outage narrative is deliberately not copied into the runbook. Repo
convention for the ceilinged instruction surfaces is that provenance is one line
and stories live on cards, not in operational text — so the story stays here and
on the card's addendum, and the doc carries the rule it produced.

Ceilings — paid by consolidation, nothing raised

file	before	after	ceiling
.claude/skills/pm-dispatch/SKILL.md	1008	1005	lowered 1008 → 1005, headroom 0
.claude/skills/pm-dispatch/references/dispatch-runbook.md	274	274	274, headroom 0

The superseded prose is the purse: the 3–5 budget clause, the >15 trigger
sentence and the 15–20 batch parameter funded the ≤5 convention and the alarm
line. scripts/pm/check-skill-line-ratchet.mjs moves only to lower the
SKILL.md ceiling to the landed count — the shrink-only direction the ratchet
calls always legitimate. No registry entry was raised.

Every line stays within MAX_LINE_BYTES=120; the two verbatim rulings are
byte-intact, and no quote already spanning lines was re-flowed (the 出处
paragraph and the 状态转换机制 bullet are kept byte-identical).

Checks — all green at a879b1e57

Derived with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack
(14 families) plus check:nul-bytes:

check:agent-test-spelling · check:cross-package-test-inputs ·
check:doc-authoring · check:doc-formula-expressions · check:entry-guard ·
check:parse-guard · check:pm-governed-merges · check:pm-governed-prose ·
check:pm-skill-id-lint · check:pm-skill-ratchet · check:pnpm-filter-targets ·
check:skill-frame-sync · check-ci-filter-parity.mjs ·
check-cross-package-test-inputs.mjs · check:nul-bytes — 15/15 exit 0.

Verdict lines: ✓ check-skill-line-ratchet: …SKILL.md is 1005 lines (ceiling 1005; headroom 0). · ✓ check-skill-line-ratchet: …dispatch-runbook.md is 274 lines (ceiling 274; headroom 0). · ✓ check-skill-line-ratchet self-test: 71 cases pass.
· ✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files · check-nul-bytes: OK (scanned 6658 text file(s) …; no raw ASCII control bytes).

No changeset: .claude/** + scripts/pm/** only — nothing published ships from
this diff.

---

Generated by Claude Code