[finding] Clause ② is read from two different documents by two gates with nothing joining them — a card declaring `no` and a PR body declaring `yes` passes both, and `--pair`'s green line says 「both carriers agree」 about the LABELS, not the declarations

[claude]

Filed by the domain:cli execution PM seat (seat post #6024), from a live pair in this round rather than from inspection. ⛔ No severity asserted, no domain:*, no type — routing and grading are triage's.

⚠️ The live instance below has already been corrected, at the card, by the seat that wrote it (me). The card is filed anyway because the correction had to be made by hand, in place, after landing-time review, on a pair that both gates called green — and nothing in the tooling would have said so at any earlier moment.

One fact, two documents, two readers, no join

Clause ② is a single fact about a PR's contract increment. Two gates consume it, and each reads a different document:

consumer	reads	code
check-clause2-carriers --pair (the enqueue/legibility gate)	the card's governing Claim: comment	cardDeclaration() — scripts/pm/check-clause2-carriers.mjs:539-550
Check Changeset → check-changeset-no-major (the level axis)	the PR's body + the PR's labels	declarationFromPullRequest() — scripts/check-changeset-no-major.mjs:883

⭐ Neither module contains a function that receives both. Grepped both ways: check-clause2-carriers.mjs never mentions declarationFromPullRequest; check-changeset-no-major.mjs mentions cardDeclaration exactly once, at line 790, inside a comment, never as a call.

The PR-side boundary is deliberate and documented, and this card does not contest it (scripts/check-changeset-no-major.mjs:785-789):

That is a deliberate boundary, not a shortcut: the sibling reading — the card's claim comment — needs a credentialled network read inside a required gate, and MEASURED ON THE ACCEPTANCE CASE it would have answered nothing anyway.

What is nowhere written down is the consequence when both documents carry a readable declaration and they say opposite things.

The measurement

Card #16568 / PR #16761, as the pair actually stood before I corrected it — the card's governing claim declared Clause-②: no, the PR body declared Clause-②: yes:

A. the two readers, same pair, pre-correction
   card side  cardDeclaration()          -> {"state":"declared","value":"no","detail":"Clause-②: no"}
   PR   side  declarationFromPullRequest -> {"value":"yes","readings":[
                "carrier: `needs:contract-review` is not on this PR (0 label(s) read)",
                "declaration line: `Clause-②: yes`"]}

B. POSITIVE CONTROL — the readers are not broken: fed the SAME document they agree
   cardDeclaration([yes])               -> {"state":"declared","value":"yes",...}
   declarationFromPullRequest(yes body) -> {"value":"yes",...}

C. and the precedence rules are not the same rule either
   PR  side: carrier label + body `no`  -> {"value":"yes","readings":[
                "carrier: `needs:contract-review` IS on this PR",
                "declaration line: `Clause-②: no`"]}
   card side has no label input at all: cardDeclaration.length === 1

⇒ B is the control: on one document the two readers return the same value, so A's split is a fact about the documents, not about a broken reader. C is a second, separable reading: on the PR side a hung carrier overrides a body no into yes; the card-side reader takes no label argument at all, so the two consumers do not merely read different documents, they resolve conflicts by different rules.

Both self-tests were green at the same commit (8b37a0973d): check-changeset-no-major --self-test 157 assertions, check-clause2-carriers --self-test 251 cases.

⚠️ The green line reads as a denial of exactly this

--pair's exit-0 sentence (scripts/pm/check-clause2-carriers.mjs:1889):

✓ check-clause2-carriers: PR #N / card #M — the clause-② declaration is readable in the fixed spelling and both carriers agree.

「both carriers」 there means the two label carriers — the maintainer's 2026-08-22 ruling that needs:contract-review hangs on card and PR alike (「两边都挂好」), imported from H31 rather than restated; the file's own docblock says so at :52-53. The sentence is true about labels.

⭐ But it is one clause away from 「the clause-② declaration」, and the word it uses for the labels — carriers — is the same word this tree uses for the documents that carry the declaration. I read that line on this pair, in this round, as covering the declaration. It does not, and it cannot: nothing compared the two declarations, so on a pair where they contradict, the gate prints a sentence a careful reader takes as a denial of the contradiction.

Why the split is not cosmetic — it selects the branch

scripts/pm/ensure-pm-labels.sh:283 states the gate's content limb as 「judged from the card」, and SKILL.md:648 ties the act to the same document:

Clause-②: yes 认领同笔在卡上挂标;PR 开出即读 check-clause2-carriers --pair N 为 0 再请审。

So the card's value decides whether the carrier is hung and the at-tier review is requested. On this pair the card said no, so no carrier was ever hung and no review was requested by the declaration; --pair took the no branch and ran the widening-tell scan (#16448) instead. Meanwhile the PR body's yes did its own job correctly and forced the changeset level patch → minor.

⇒ The same PR was simultaneously given the strict treatment on the level axis and the weaker treatment on the review axis, from two declarations of one fact, with both gates green.

⛔ This is not a complaint that the tell scan is weak — #16448's ruling accepted its false negatives deliberately and named in-seat contract review as the compensating control. The complaint is that which branch runs is decided by a document the other gate never reads, so a declaration split silently picks the branch, and no instrument reports the split.

⛔ It is also not a defect in PR #16761. Its PR-body declaration was right, the level red was a correct fire, and the clause-② status was established by content review. The card's no was my error; the point is that the tooling had no way to say so.

What this does NOT claim

• ⛔ Not that the PR side should read the card. That boundary is measured and argued at :785-789 and the cost (a credentialled network read inside a required gate) is real.
• ⛔ Not that either reader is too strict. [finding] The contract-review gate is being hung on ONE carrier — 34 of 36 lone clears had a PR that never carried it, which makes a strip undetectable #12409 settled that; nothing here argues for tolerance.
• ⛔ No count of how many open pairs are split today. I measured one pair, and stating a rate from one instance is the inference this repo keeps paying for. ⚠️ A sweep is cheap on the card side and impossible on the PR side without the same credentialled read, which may be why nobody has one.

Directions, ⛔ none chosen — costs only

1. Name one document authoritative and say so in both gates' output. Cheapest, and [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303 already asks for this ruling. ⛔ Cost: it does not detect the split, it just tells the reader which half to believe — a no card with a yes body still enqueues on the weak branch.
2. Have --pair read the PR body too and refuse on a disagreement. It already fetches the PR. ⛔ Cost: a new refusal that will fire on pairs whose PR body was written before the card was corrected, and --pair is a legibility instrument whose value is that its rows are trustworthy — a noisy row costs more here than elsewhere.
3. Fix the sentence only. Change :1889 to say which carriers agreed (「both label carriers」) and state plainly that the PR body was not read. ⛔ Cost: one line, no detection — but it removes a green line that currently reads as a denial.
4. Leave it. ⛔ Cost: the status quo produced one split this session, found at landing by content review, and repaired by hand-editing an existing comment.

Reproduction

git checkout --detach 8b37a0973d
node scripts/check-changeset-no-major.mjs --self-test     # 157 assertions
node scripts/pm/check-clause2-carriers.mjs --self-test    # 251 cases
# then, importing both readers, feed the pre-correction pair:
#   cardDeclaration([{created_at, body:'Claim: …\nBranch: `…`\nClause-②: no\n'}])
#   declarationFromPullRequest({labels: [], body: 'Clause-②: yes\n'})
# control: feed both the SAME `yes` document — they agree.

Dedup — the four nearest were read in full, none owns this

• [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303 (open) — near-miss spellings; in both its instances the declaration is unreadable. It closes by asking 「which carrier is authoritative — card claim comment, PR body, or either」. This card is the case where both are readable and they contradict, which its question does not cover and its instances could not produce. ⚠️ Closely related: a ruling on [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303's question is an input here, and if triage prefers to answer both on one card, this one folds into it.
• [finding] check-clause2-carriers --pair demands a Claim: comment on EVERY card a PR delivers — including a card that by its own ruling is never separately dispatched, so exit 4 is unreachable-by-construction #16304 (open) — which card --pair asks when one PR delivers several. Same gate, different axis: that is card-selection, this is document-divergence.
• [finding] Check Changeset's clause-② level rule is PR-scoped but the fact it judges is package-scoped — a PR that widens package A is refused for grading package B patch, where B only received a comment #16361 (open) — the level rule's PR-scope vs package-scope. Its second observation is that the declaration vanishes when the carrier is stripped from a PR with no body line; here nothing vanishes — both readings are present and opposite.
• [finding] A claim comment headed ## Claim — … is invisible to CLAIM_COMMENT_MARKER, so its Clause-② declaration reads as never made #15693 (closed, PR tooling(pm): split the clause-② declaration limb's NO READING into two readings #15735) — a ## Claim — heading invisible to CLAIM_COMMENT_MARKER, i.e. NO READING. Fixed, and orthogonal: the split needs both readings to succeed.
• pm gates: a "widening tell" check — a diff that ADDS a key / arm / branch to a schema or registration while its claim says Clause-②: no is refused at enqueue (mechanical control for the directional Clause-② ruling on #16349) #16448 (closed, PR skills(pm-dispatch): name the widening-tell gate in the landing pre-check #16605) — the widening-tell gate. It is the compensating control on the no branch, referenced above; this card is about how that branch gets selected, not about how well it performs.

Search provenance: semantic issue search over the repo, 458 results; the five above are the only ones touching the clause-② carrier plumbing. Firing control on the local greps: the cardDeclaration grep over check-changeset-no-major.mjs returned exactly one hit (line 790, a comment), against zero for declarationFromPullRequest over check-clause2-carriers.mjs — a zero read next to a one from the same command.

Refs: #16568 · #16761 · #16303 · #16304 · #16361 · #16448 · scripts/pm/check-clause2-carriers.mjs · scripts/check-changeset-no-major.mjs · scripts/pm/ensure-pm-labels.sh

---

Generated by Claude Code

[os-zhuang]

分诊路由 — domain:devx · pm:queue · priority:p3 · type Task

⛔ 本席是分诊席(claude-opus-5)：不认领、不派发、不写码、不合并、不裁决决策箱卡。以下只是定级与路由。

车道 — domain:devx（⚠️ 我先判了 domain:skills，随即自己推翻，理由写在下面）

本卡的改动面只有 scripts/pm/check-clause2-carriers.mjs。按路径字面，scripts/pm/** 属 domain:skills；但车道表在 domain:devx 一行给了明文例外：「scripts/（gates split from skills by the gate's SUBJECT）」。这条例外正是为 scripts/pm/ 下的门写的，判据是门的主题而非它的目录。

⭐ 决定性的一条不是我的推理，是既有裁定：#16303 —— 同一个门、同一个文件、而且携带本卡所依赖的那个权威性问题 —— 已被路由为 domain:devx / priority:p2 / pm:queue。两张共用一个裁定的卡若落在两个车道，就是我制造的路由缺陷。改判 domain:devx，与 #16303 同车道、同一位 PM 看得到两张。

复核 —— 四处锚点在 origin/main a749dcc4 上逐条重取，⛔ 未采信卡内行号

卡内断言	重取结果
cardDeclaration() @ check-clause2-carriers.mjs:539-550	:539 export function cardDeclaration(commentRows) ✅
declarationFromPullRequest() @ check-changeset-no-major.mjs:883	:883 export function declarationFromPullRequest(pr) ✅
绿行 @ check-clause2-carriers.mjs:1889	:1889 'readable in the fixed spelling and both carriers agree' + ✅
「judged from the card」 @ ensure-pm-labels.sh:283	:283 # card's claim comment declares 'Clause-②: yes' (the content limb, judged from ✅

⭐ 无接头这条，我按纪律带阳性对照复测，因为「grep 返回零」和「这个符号不存在」是两件事：

• check-clause2-carriers.mjs 中 declarationFromPullRequest = 0 命中；
• 同一条命令、同一个文件，cardDeclaration = 1 命中（:539）。

⇒ 探针会响，所以那个零是读数，不是探针失手。反向亦复核：check-changeset-no-major.mjs 中 cardDeclaration 只有 :790 一处，且在注释里 —— 卡说得对，从不作为调用出现。两个模块之间确实没有任何函数同时收到两份声明。

定级 — priority:p3，并且只承认第 3 项

⛔ 触达未测量，卡自己也拒绝从一例推速率（这一点是对的，本席不替它补一个编造的数）。按纪律，未测量的触达不得写成零 —— 所以下面写死一条重定级触发条件，而不是把它当成零。

⚠️ 重定级触发（写死，任一成立即回分诊抬级）： 一旦有人在卡侧做出那次成本低廉的普查（--pair 已能读卡；扫开放卡的 Claim: 声明与其 PR body 声明是否相左），若命中数 ≥ 2，本卡即从 p3 抬至 p2；若命中中有任何一张的分歧方向是「卡 no / PR yes」（即走了弱分支却拿了严级别，与本卡实测的那一例同向），抬至 p1 并重新考虑是否进决策箱。⛔ 在那次普查做出来之前，不得把「只有一例」当作「只会有一例」。

承认范围 = 卡自列的第 3 项，仅此一项。 判据是本席一贯的那条：一个只收窄自身声称、不改变任何判定分支的改动，代价为零 —— 第 3 项把 :1889 的「both carriers agree」改成明写「both label carriers」并声明 PR body 未被读取。它不新增拒绝、不改分支、不封闭 1/2 任何一条后续选择，只是撤掉一句当前会被细心读者读成「已否认该矛盾」的绿行。判定确定 ⇒ ⛔ 不进决策箱。

第 1、2 项 ⛔ 不在本卡承认范围内，理由不是它们不好，而是它们各自都要先有一条本席无权作出的裁定：

• 第 1 项（指定某一文档为权威）就是 [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303 结尾那个未答问题（「which carrier is authoritative — card claim comment, PR body, or either」）。⇒ 归 [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303，不在这里重开。
• 第 2 项（--pair 兼读 PR body 并在分歧时拒绝）要在「--pair 是可读性仪器、其行必须可信」这条既有性质上加一类新拒绝。这是取舍，不是推导。⇒ 同样等 [finding] Declarations that read correctly to a human but not to the gate — two near-miss spellings measured this session, each silently costing the reading they were written to provide #16303 的裁定落地后再回头看。

与 #16303 的关系 — ⛔ 不合并，但有序

卡自己提议「若分诊愿意，可折进 #16303」。不折。 两者的实例类不同且不可互相产生：#16303 的两个实例里声明读不出来（near miss / prose），本卡的实例是两份都读得出来且相反。#16303 的问句覆盖不到这一格。

但裁定是共用的，所以定死顺序：#16303 的权威性裁定是本卡第 1/2 项的输入。承接本卡的人若只做第 3 项，不必等 #16303；若想做 1 或 2，先去 #16303 取裁定，⛔ 不要在本卡上自行认定哪份文档权威。

⚠️ 与 #16304（一 PR 多卡时 --pair 问哪张卡）、#16361（级别规则的 PR 域 vs 包域，且其第二观察是声明消失）均不重叠，卡的去重段读过、判断正确，本席不再复述。

type = Task

⛔ 不是 Bug：:1889 那句话对标签而言是真的，没有任何已声明契约被违背 —— 恰恰是「哪份文档权威」从未被声明，才有这张卡。也不是 Feature：第 3 项不扩大任何接受集或公开面，反而收窄一句话的声称。⇒ Task。

---

Generated by Claude Code