skills(pm-dispatch): a ruling's file/symbol/line/call-shape clause carries its reading or is marked 未验证 — the 派发令 evidentiary duty extended to 裁决 - #18679

Fixes #18506

Clause-②: no

skip-changeset — the only touched path is .claude/skills/pm-dispatch/SKILL.md (governed .claude/**, internal agent tooling; it ships in no package's files[]).

The defect, in one paragraph

A ruling is consumed exactly like a dispatch order — quoted verbatim into the brief, ⛔ not re-litigable, treated by the dev as settled — yet the charter's evidentiary discipline binds dispatch orders only: 「派发令里关于代码的危害断言必须有读数(点名 call site / 路径 / 迁移)」, 「搬自任何工件(含己文)本轮未亲测计数/零命中/文件面:重测或注「未验证」点名来源」, 「PM 测不了的危害 ⇒ 写成给 dev 的问题,⛔ 不写成栅栏」. Nothing required a ruling's code-level assertions to carry a reading, and nothing marked them unverified when they did not. The artefact with more authority had fewer evidentiary requirements than the one that merely relays it — the asymmetry is the defect. This PR extends the same two options (carry the reading, or mark that clause 「未验证」) to 裁决, so the dev reading a ruling knows which clause to falsify first. ⛔ It is not 「measure before ruling」, adds no gate or script, and re-grades no ruling.

The evidence: six for six on one split (card comment 5701534399)

The rulings that reasoned from rules held; the four that asserted implementation facts broke. The exposure is confined to the moment a ruling names a file, a symbol, a line or a call shape — the same boundary the charter already draws for dispatch orders, which is why the line below carries a rule-citing exemption rather than a measurement duty.

The change — one line in, one line out, SKILL.md 812 / 812

In — .claude/skills/pm-dispatch/SKILL.md :731 (work tree), section 升级与决策, between 「卡上裁决评论是详注,仍是执行与复核的第一落点。」 and 「出决策箱须引同趟取回过的裁决 id …」:

- 裁决点名文件/符号/行/调用形状的条款,逐条带读数或注「未验证」;只引规则的条款免。

118 B (cap 120). Reads: a ruling clause that names a file, a symbol, a line or a call shape carries, per clause, the reading behind it or is marked 「未验证」; a clause that only cites rules owes nothing. 读数 is the corpus's defined term (:166–:167: UTC 取数时刻 + ref/tip; AGENTS.md: a count plus the tree it was taken against), and 「未验证」 is the dispatch-order rule's own mark, so the mark travels verbatim with the clause when the ruling is quoted into a 派发令.

Out — the 分诊座位职责 line at origin/main :309, 95 B:

- 代裁只由维护者召唤的总监席产出,⛔ 永不凭自述、⛔ 无子代理裁决。

Every clause of it is already stated in the ruling section this PR adds to: 「总监席是唯一裁决者:决策箱、代裁与一类自裁只出自本席,⛔ 无子代理无自述。」 (:769) covers 代裁 ⊂ {决策箱、代裁、一类自裁}, 只出自本席, 无自述, 无子代理; 「项目总监席:维护者在所在仓召唤 …」 (:765) covers 维护者召唤; :75 「裁决不在本席,归维护者召唤的总监席」 and the red line :30 「唯一例外是已裁的代裁通道」 still tell the triage seat the same thing in its own sections. Deleting the second spelling is the corpus discipline (one rule per line, no rule stated twice), not a weakening — a control grep for the deleted spelling returns 0 and each restating line returns its line.

Why no cheaper payment existed: measured on this tree, of SKILL.md's 598 adjacent bullet pairs ZERO merge under the 120-byte cap (smallest 124 B), and in the 升级与决策 section ZERO of 40 (smallest 158 B); re-wrap funding is refused by the ratchet rule. The whole-file diff vs origin/main is exactly these two hunks (1 insertion(+), 1 deletion(-)).

Placement — SKILL.md, not core-rules.md, along the four axes

• 实际业务需求 — the consumers are the seat writing a ruling and the dev reading it quoted into a brief; both read SKILL.md in full every session. core-rules.md :3 says 「本文不新增规则」 and :43 makes it 「本文的子集」, so it cannot be the sole home; a twin there would cost a second line in a file at 151 / 151 where ZERO of 122 adjacent pairs merge under the cap (smallest 175 B, measured on this tree) and no line restates another — an unrelated compression to fund it, which the ratchet's own precedent refuses.
• 项目长远合理性 — the line sits in the block that defines the ruling record (:729 权威载体, :730 卡上裁决评论, :732 出决策箱须引裁决 id), where 裁决 is the generic term — maintainer ruling, 代裁, 一类自裁, PM 裁量, retriage 改判 — so it binds every seat that rules without naming one. :44 (一条规则在本文与核心条款一处改动,另一处同 PR 同改) is not triggered: the new line is new, and the three dispatch-order lines whose twin is core-rules :116 are byte-untouched.
• 防 AI 写错 — the mark is mechanical (a clause either has a 读数 or the literal 「未验证」), per clause, and rides verbatim into the 派发令 under the existing 「裁决不可重裁 / 机制假设须实测」 partition; no seat has to judge whether it is a nice-to-have, and a 「不 expected」 built on an unread function is no longer a legal spelling.
• 创业阶段不扩散 — one line, zero net lines, no gate, no script, no new label, no re-grade; the rule-citing exemption keeps the cheap reading and forecloses the expensive one.

Measured: the neighbours do not contradict it, and the quoted lines are untouched

Neighbouring text (work tree): :715 「② 把裁决挂在具名、可证伪的前提上,派发令要求 dev 先验前提再动手。」 (a premise-bearing ruling names its premise — compatible: the new line says which clauses of any ruling are the premises); :180 「裁决明令的动作实施中测出对向事实 ⇒ 照字面执行 …」 (what happens when a measurement contradicts a ruled action — unchanged); :427–:432 (the PM's pre-dispatch staleness re-read of a ruling's 动作面/卡引用面 — about the re-read, not about what the ruling carries); :401 「① 既有裁定/在案类规则/成文纪律机械决定方向,裁决评论点名所用权威」 (the class-1 gate is rule-citing by construction — the exemption agrees with it); references/lanes/director.md :11 (the ruling comment's fixed first line — unaffected).

Byte-identity against origin/main 7f7b8557d (diff of the extracted lines, empty): the three quoted dispatch-order lines (:541–:543 there, :540–:542 here); PR #18656's :91; PR #18666's :198–:209, :536–:542 (now :535–:541), :625–:631 (now :624–:630), :642–:648 (now :641–:647). The set of over-120 B lines is unchanged (23, all table rows plus the claim-template line); widest row 342 B, pin 342. Line numbers shift by −1 between the deleted line and the inserted one; no hunk overlaps either PR's hunks.

Readings

Head cb1a949d0 (origin/main f6c2eb7c8 merged — one unrelated driver-sql commit; the SKILL.md delta unchanged). Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack from the worktree (change set from the merge base, no hand-fed paths): 18 commands, list identical before and after the merge. Exit codes captured redirect-then-$?, reconciled with --ran: 「18 derived, 18 run, 0 NOT-MEASURED (a DERIVED zero — all 18 recorded an exit code and none of them is 3)」.

node scripts/pm/check-governed-merges.mjs --test .claude/skills/pm-dispatch/SKILL.md → exit 3, 「GOVERNED — a human merge is the review record for this PR」 — recorded as the expected answer, not a failure. CI-owned and not run locally: the 52 artifact-roster families, the 11 wide-population families and pnpm lint (the derivation names them as outside the 18; none reads this file's content). No package build is owed by the diff (no package touched); the one build above was that gate's own prerequisite.

Acceptance notes

• noted, not filed: references/lanes/director.md :9 「裁决只存在于本卡:聊天原话、台账行、他卡引述 ⛔ 都不是裁决」 and SKILL.md :729 「一条裁决已记录的权威载体 = 总监席每场收工的摘要台账」 name different carriers as authoritative for a ruling. Observed only; not a defect class (a)–(c), and outside this card. 承接者:无.
• noted, not filed: core-rules.md carries no twin of the new line (reasons above). If the maintainer wants one, it is one core-rules line at 151 / 151 with no duplicate to pay — a ruled raise or a compression of another rule, both the maintainer's call. 承接者:the domain:skills seat, at the maintainer's word.

维护者速读(草稿)

改了什么 — pm-dispatch 技能文件加一行、删一行,行数不变(812 / 812):裁决里凡点名文件、符号、行号或调用形状的条款,要么带上得出它的读数,要么该条注「未验证」;只引规则的条款不欠。删掉的是分诊节里一条与总监席节完全重复的「代裁只出自总监席」拼写。

为什么改 — 派发令早有这条纪律(代码危害断言必须有读数,未亲测就注「未验证」),但裁决没有。裁决被 dev 当作已定事实照抄进派发令,不可重裁。一轮六张卡实测:引规则的两条裁决全对,断言实现事实的四条全错(文件位置、调用形状、挂载拓扑、折叠方向),每次都靠 dev 自己重新测量才发现。权威更高的工件反而举证要求更低,这是结构性缺口,不是三次失误。

风险与代价(含回滚) — 零新增门禁、零脚本、零标签;只约束裁决的写法,不要求先测再裁(引规则免费,只有点名代码事实的条款才欠读数或标记)。代价是出裁决的席位多写几个字。回滚 = revert 这一个 commit,两行原样回来。

席位意见 — (留空,由席位填写)

你要做的(一个动作) — 批准这个 PR(受管面 .claude/**,需要你的 approve 或亲手合并)。

Generated by Claude Code