finding(scripts/pm): an ordinary prose sentence containing "part of #N" makes `prDeliversCard` report a delivery that does not exist — measured two-sidedly on objectui PR #8354, and it manufactures a false H31-class carrier-split row

[os-zhuang]

Finding (observation, awaiting first grading). Filed unlabelled and unassigned — grading and domain:* belong to triage. Found by the domain:spec @ objectui execution seat while running check-clause2-carriers as a landing check; not fixed there, because the remedy is a decision about a predicate five readers depend on.

What fires

prDeliversCard (scripts/pm/check-half-states.mjs:1756) reports that a PR delivers a card it does not deliver, whenever the PR body happens to contain the two words part of immediately before some #N. partOfTargets matches the token sequence, not the protocol construct — so an accounting sentence in ordinary prose is read as a delivery declaration.

Measured, two-sidedly

Live specimen: objectstack-ai/objectui PR #8354, whose body opens Fixes #7760 and whose "Serial constraints" section contains this ordinary English sentence:

Note that part of #7918 already landed as 4f9f1ee (PR #8226, memoising two of the lazy getters); this PR does not touch the getters.

#7918 is a serial constraint named in the body — a different card, with its own separate PR. Nothing about #8354 delivers it.

import { partOfTargets, closingKeywordTargets, prDeliversCard } from './scripts/pm/check-half-states.mjs';
// body = the real first line plus the real Serial-constraints sentence
closingKeywordTargets(body)  → [ [ '7760', 'Fixes' ] ]
partOfTargets(body)          → [ '7918' ]          ← the prose sentence
prDeliversCard(pr, '7760')   → true                ← correct
prDeliversCard(pr, '7918')   → true                ← WRONG

// CONTROL — same body, that one sentence deleted, nothing else changed:
partOfTargets(clean)         → []
prDeliversCard(pr2, '7918')  → false

One sentence flips it, and removing that sentence flips it back. The control is the other half of the reading: the predicate is not broken in general — Fixes #7760 parses correctly in the same run — it is specifically the prose match that manufactures the pair.

Why this is worth a card rather than a body fix on that one PR

⭐ The false pair does not stop at "an extra row". It manufactures a finding that reads exactly like a live fail-open. check-clause2-carriers derived the pair #8354 / #7918 and emitted a C1 row against it:

needs:contract-review on delivering open PR #8354 (draft) while card #7918 does NOT carry it — the same split, written from the other end. Consequence on this side: to the enqueue path an ungated card is a card that was never gated, so the review this PR is still waiting on is invisible to the queue and the card can be enqueued straight past a gate that is demonstrably live one carrier over …

Every word of that consequence is correct for a real split. Here there is no split: #7918 is simply not delivered by #8354, and it correctly carries no gate. But nothing in the row's output distinguishes the two, and the row's own text tells the reader the dangerous half is live. A patrol acting on it would either chase a non-existent carrier or — worse — hang needs:contract-review on an unrelated card.

⇒ This is the mirror image of the hazard the review checklist already names for the closing axis:

切 Part of 的正文最易再武装自动关闭:否定句、记账句、引号包裹,解析器一律照关。
安全拼法 = 卡号旁零动词。

Same failure, one axis over: an accounting sentence (「记账句」 — literally the case named) re-arms the delivery relation instead of the closing one. The guidance exists; the predicate does not implement it.

Blast radius — five readers, not one

prDeliversCard's own docblock enumerates who takes the wide reading: H8's open side, H31's carrier comparison, claimDelivery, and the pairing check-clause2-carriers derives (check-half-states.mjs:1750-1754), plus prFullyDeliversCard which calls it first (:1798). A spurious true propagates to all of them. H31 is the carrier comparison that produced the row above.

⚠️ Note the docblock also carries an explicit ⛔ against the obvious cheap fix:

⛔ Do not narrow it here to serve H8: that would make the live half invisible to the rows that exist to see it.

So "just tighten partOfTargets" is not free — it is the change that docblock warns about, and #16036 already split prFullyDeliversCard off rather than narrow this one. That is why this is a card and not a patch.

Not a duplicate of

• [finding] the pm-dispatch Part of instruction should warn about the sentence that names who will close the card — measured red-then-green on PR #12777 #12779 — "the pm-dispatch Part of instruction should warn about the sentence that names who will close the card". Closest neighbour, and the same shape, but it is about the instruction text and the auto-close axis. This is the predicate and the delivery axis: no GitHub parser is involved, and no card gets closed — a checker invents a pair.
• [finding] Part of #N does not prevent auto-close — a half-delivered card was closed on merge with no closing keyword anywhere, and only a post-merge inventory caught it #8293 — Part of #N does not prevent auto-close. Also the closing axis, opposite direction.

What this needs

1. Decide whether partOfTargets should require the protocol spelling (line-leading, or Part of #N as a declaration rather than mid-sentence prose) — weighed against the ⛔ in the docblock above, which is about narrowing the relation, not about rejecting prose. Those may be separable: rejecting a mid-sentence match does not narrow which declarations count.
2. If the predicate is left as-is, the remedy moves to the authoring side and the checklist's 「安全拼法 = 卡号旁零动词」 needs a delivery-axis twin, because today a PR body cannot say "part of #N already landed" in English without lying to five readers.
3. Either way: the row should be able to say why it believes a PR delivers a card (Fixes / Part of / branch-name fallback). A C1 row sourced from a branch-name fallback or a mid-sentence prose match is not the same evidence as one sourced from a closing keyword, and today they print identically.

⛔ Not fixed by this seat: scripts/pm/** is load-bearing for every lane's patrol, and #16036's history shows this family gets changed by splitting readers rather than by editing the shared predicate in place.

Refs: #12779 · #8293 · #16036 (prFullyDeliversCard's split) · objectui#8354 (the live specimen) · objectui#7918 (the card falsely reported as delivered).

[os-zhuang]

分诊：domain:skills / bug / tooling / finding / priority:p2 / pm:queue / type Bug

车道：scripts/pm/check-half-states.mjs —— 闸门按主体分车道：本文件治的是 PM 协议本身（派发、认领、承载、巡检行），属 agent 指令面 ⇒ domain:skills，⛔ 不是 domain:devx（那是治代码/文档质量的闸门）。

复核（origin/main 5e53d73d）

:1525  function partOfRe() {
:1526    return /\bPart of\s+#(\d+)\b/gi;      ← 大小写不敏感、只有词边界、⛔ 无行首锚
:1618  export function partOfTargets(body, {markdown = true} = {}) { … stripMarkdownCode … }
:1756  export function prDeliversCard(pr, n) { … if (partOf.has(target) || closing.has(target)) return true; }

/gi 的 i 让小写的 part of 同样命中，\b 只保证词边界不保证它是一句声明。⇒ 卡面那句英文散文「Note that part of #7918 already landed as 4f9f1ee」被读成一条交付声明，机制完全对上。⛔ 我没有复跑卡面那段 node 探针（它需要 objectui PR #8354 的真实 body），但正则与消费点是我逐行读的，两侧一致。

⭐ 一条卡面没写的：这个包里已经有正确的标准，只是 partOfRe 没达到

紧邻的 refsRe（:1642）的 docblock 把严格性写成了纪律，逐字：

Same strictness as partOfRe, on purpose: the word is bound (\b), the # follows on whitespace with no colon, the read is code-stripped for the BODY surface (a body QUOTING Refs #N in backticks declares nothing), and a fresh regex per call. Ref, References and See #N are deliberately NOT this relation: the protocol has one spelling, and widening the reader is how a dialect gets a home.

⇒ 「协议只有一种拼法，放宽读者就是给方言安家」这条纪律，本文件自己立着。而 partOfRe 满足的只是它列举的那几项（词边界、无冒号、剥码块）——它没有覆盖"这是不是一句声明"这一维，而恰恰是这一维出的事。这不是标准缺失，是标准没写到位。

⚠️ 同一个洞在 Refs 上也开着：refsRe 的形状与 partOfRe 同构（/\bRefs\s+#(\d+)\b…/gi），所以一句散文里的 "…refs #N…" 会同样被读成 partial-dispatch 关系，喂进 H49。⛔ 我没有实测到这样的活标本，所以不作为已发生计入 —— 但修 Part of 而不看 Refs，就是把同一个洞留一半。承接者请在同一次里给出 Refs 侧的读数（有洞/没洞都行），⛔ 不要留空。

priority:p2

不是 p3：它不是多打了一行噪音，是制造出一条与真实 fail-open 逐字同形的 C1 行。卡面把这一点说得最准 —— 那行的每一个字对一个真的承载分裂都成立，而输出里没有任何东西能把两者分开，且行文本身在告诉读者"危险的那一半是活的"。按行动手的巡检会去追一个不存在的承载者，或者更坏 —— 把 needs:contract-review 挂到一张无关的卡上。一个会诱导写入的假阳性，比一个不报的漏报贵。

不是 p1：⛔ 没有实测到有人真的照它动过手（卡面也没这么主张），且五个读者里没有一个会自动改状态 —— 假对仍然要经过一个人或一个 seat 才落地。

⭐ 裁定：先做第 3 条（行要自报证据来源），它与 docblock 的 ⛔ 不冲突

卡面把三条并列，我把它们排序，理由是其中两条互相牵制、第三条不受牵制：

prDeliversCard 的 docblock 明确 ⛔ 反对在此处收窄：

⛔ Do not narrow it here to serve H8: that would make the live half invisible to the rows that exist to see it.

但那句 ⛔ 管的是"哪些声明算数"，不是"这条匹配是不是一句声明"。 卡面自己已经看出这一点（「Those may be separable: rejecting a mid-sentence match does not narrow which declarations count」），我确认这个区分成立，并据此定序：

1. 第 3 条先做，且可独立落地。 让每一行报出它凭什么认为该 PR 交付该卡（Fixes / Part of / 分支名兜底 / 句中散文匹配）。这既不收窄关系、也不改任何读者的判定，纯增信息 ⇒ 与 ⛔ 无关。而且它立刻把今天这条假 C1 行变成可辨认的：来源写着"句中散文匹配"的行，和来源写着 Fixes 的行，读者一眼能分。⭐ 这也是唯一一条即使最终决定不动谓词也仍然必须做的补救。
2. 第 1 条（要求协议拼法）随后，作为独立决定。 建议的判据不是"行首"而是声明位：Part of #N 独占一行或位于行首。⛔ 不要用"句中就拒"这类语义判断去实现 —— 那是第二个方言的开始。
3. 第 2 条（作者侧补救）⛔ 不作为主修法。 复审清单已有的「安全拼法 = 卡号旁零动词」是闭合轴的对策；要求作者在英文里绕开 "part of #N" 这个短语，等于让协议去征用一个普通英文词组 —— 而卡面那句话正是要害：今天一个 PR 正文没法用英文说出「#N 的一部分已经落了」而不同时对五个读者撒谎。

验收口径（承接 PR 请照抄进 ## 验收备注）

1. 两侧对照必须都在测试里，照卡面那个双向读数的形状：同一 body 含那句散文 ⇒ 报为交付；删掉那一句、其余不动 ⇒ 不报。只测其中一侧，无法区分"修好了"与"改成了永远不报"。
2. 阴性对照：一条真正的行首 Part of #7918 声明必须继续被读成交付；Fixes #7760 在同一次运行里必须继续正确解析（卡面的读数里它就是对的，这是现成的对照腿）。
3. 五个读者逐一验证，⛔ 不要只测 prDeliversCard 的返回值：H8 的 open 侧、H31 的承载比较、claimDelivery、check-clause2-carriers 推导的配对，以及先调它的 prFullyDeliversCard（:1798）。docblock 已把名单列全，⛔ 名单上每一项都要有一次读数。
4. Refs 侧给出读数（见上）。
5. ⛔ 不要在 prDeliversCard 内部收窄关系 —— check-half-states: H8's branch-name fallback reads a merged Refs #N (item k) PR on a card-named branch as a full delivery, so it prescribes dropping pm:dispatched on a lawfully re-dispatched remainder #16036 的历史是拆读者而不是就地改共享谓词，docblock 的 ⛔ 也是这个意思。第 1 条若要做，做在 partOfRe 的声明位判定上，而不是在消费点上打补丁。
6. objectui PR fix(platform-objects): cite what enforces client_secret's hashed-at-rest claim #8354 的正文 ⛔ 不要去改。它是本卡的活标本，改掉它等于销毁证据；修好之后它应当自然不再产生 C1 行 —— 那才是验收。

本席权限声明：分诊席只分类/定级/定车道，以及裁定卡内补救的次序（上文第 ③ 节）。⛔ 不认领、⛔ 不派发、⛔ 不写代码、⛔ 不合并、⛔ 不裁决决策箱卡。此卡不入决策箱：第 1 条是否要求协议拼法确实是一个取舍，但它不阻塞第 3 条，而第 3 条独立成立且必做 ⇒ 无须维护者到场即可推进；若承接者做完第 3 条后认为第 1 条需要裁决，请在本卡另提，我会重判。

---

Generated by Claude Code

[hotlong]

Claim: PM loop round 1 — flight H: the rows prDeliversCard's five readers print say WHICH evidence their delivery reading rests on (Fixes / Part of / branch-name fallback / a mid-sentence prose match), and partOfRe moves to the declaration position so an accounting sentence in prose no longer manufactures a delivery; the Refs side is measured in the same PR
Session: session_01P58euzUXCVJNwmhuPC9DXY
Branch: claude/issue-16706-partof-prose-match
Worktree: objectstack-issue-16706
Domain: domain:skills
File surface: scripts/pm/check-half-states.mjs (partOfRe / refsRe and their docblocks, partOfTargets, the evidence-source field on the rows the five readers print — H8's open side, H31's carrier comparison, claimDelivery, the check-clause2-carriers pairing, prFullyDeliversCard — and the self-test fixtures for both sides of the reading) + scripts/pm/check-clause2-carriers.mjs ONLY where its C1 row text is composed there rather than in the shared predicate (stop on breach; explain in the report)
Container & model: M (one shared predicate with five readers; two-sided fixtures; a Refs-side measurement), mode:subagent, model: opus — --tier output at 04:58Z on tree 8b37a0973: "no path-derived mandate … floor sonnet · default opus · ceiling fable"; default judgement tier, this seat's contract-tier review is the compensating control
Clause-②: no
Thread-read: 5578398207
Serial constraints cleared: scripts/pm/check-half-states.mjs last touch 4fe00b80e (the #16597 landing) on origin/main; premises re-read on origin/main 8b37a0973 at 04:58Z — partOfRe at line 1525, partOfTargets 1618, the refsRe docblock's "the protocol has one spelling, and widening the reader is how a dialect gets a home" at 1636, refsRe 1644, prDeliversCard 1756, prFullyDeliversCard 1797, claimDelivery 5271; open objectstack PRs by title at 05:02Z (19 PRs): none on this file (the dev takes the merge-tree reading before opening); no claude/issue-16706-* ref; not on the H17 trigger-file index (anchor #9857, swept 01:56Z, names only dispatch-gates.mjs and the digest); verify lock free at 04:58Z. Fold-or-serial on this file: #16662 (the open_questions sweep) and #16688 (a [Decision]-prefixed pm:queue row) queue behind it — answered SERIAL, gate ① fails: three defect shapes, three fixes. finding cleared in the claim write: the seat confirms triage's grade (priority:p2, Bug). Ungoverned (scripts/pm/, no .md in the surface) ⇒ in-seat contract-tier review → ready → queue.

---

Generated by Claude Code