fix(pm): name PLACEMENT, not spelling, when a Clause-② key sits mid-line

[claude]

Fixes #16931

Domain: domain:devx · Type: Bug · Route: A (change what the tool SAYS, not what it ACCEPTS)
Clause-②: no — diagnostic text only. The accept set, the state union, the values, the tallies and the exit register are unchanged, and that is measured below rather than argued.

The defect

check-clause2-carriers read a compact claim header like

Domain: `domain:cli` · Clause-②: no

as missing and shipped a remedy naming only spelling — so a seat was sent to hunt a typo on a line spelled exactly right. Both existing patterns anchor at start-of-line and tolerate only line-start decoration before the key, so the line was invisible twice: not read as a declaration (correct), and not quoted back as a near miss either — which is the whole job of the mechanism that exists so a near miss is not silently read as an absence.

What changed

One reporting-only detector: the fixed key followed by its colon, at a position no line-start decoration explains. The near miss now carries a reason, and the C2 row picks its remedy sentence from it.

• reason: 'inline-key' — the key is spelled exactly right and simply not at the start of a line. The row names PLACEMENT, quotes the offending line, says in as many words that the spelling is not the defect and that there is no typo to find, and sends the seat to put the line on one of its own.
• reason: 'spelling' — everything the near-miss reporter already caught. Sentence unchanged.

CLAUSE2_KEY_LINE is untouched. The standing red line is preserved verbatim in the file and in the row it still prints: 「a predicate that reads prose is a heuristic, and the measured terminus of that direction is a check that can barely fail」.

⭐ Both direction controls, measured before and after

This reader decides "is this line a declaration?" by position alone, so #16931 and #17098 are the same regex from opposite directions, and any change that loosened position here would make #17098 worse. Both directions were measured on the same harness, against origin/main c3756ff096 and against this branch.

Forward — this card's shape

	before	after
readClause2Line	null (invisible to both patterns)	{kind: 'near-miss', reason: 'inline-key', line: …}
cardDeclaration	{state: 'missing'}, no detail	{state: 'missing', detail: THE_LINE, nearMissReason: 'inline-key'}
C2 row	"the DECLARATION LINE is what is missing … the fixed spelling is …"	"the PLACEMENT of the key is what is wrong, NOT its spelling … ⛔ There is no typo to find on that line."

⭐ The state is the same in both columns. The line is exactly as unread as it was; only the sentence moved.

Reverse control — #17098's direction, ⛔ not fixed here and ⛔ not regressed

Input: a key-INITIAL bullet that merely describes, - **Clause-②: yes/Clause-②: no** — the value alone on its line, machine-read.

	before	after
readClause2Line	{kind: 'declared', value: 'yes', …}	{kind: 'declared', value: 'yes', …}
cardDeclaration	{state: 'declared', value: 'yes', …}	{state: 'declared', value: 'yes', …}

Byte-identical — the diff of the two harness runs shows no line for this control at all. Today that reading is #17098's defect; ⛔ this PR neither fixes it nor moves it. Structurally it cannot: the declaration reader returns on the first key-line match before the reporter is consulted, and the detector is written to fire only where the key is NOT at line start. That property is pinned by a case of its own (a key-initial line is never reasoned inline-key), rather than by pinning what the defective direction currently reads — pinning today's declared would put a fixture in #17098's way.

Neighbours held

Clause-②: no still declares; the #12297 reasoning-after-value allowance (Clause-②: yes — widens the accept set) still declares; the existing prose near miss still prints the same row text and only gains its reason field.

A side effect measured, not glossed

A claim comment that merely DESCRIBES with prose before the key (the :2272 fixture's shape) also carries the key mid-line, so it now takes the placement sentence too. That is why the sentence is worded to state only what is measurable — the thread carries the key, not at the start of a line — and never that the judgement was made. Its state is unchanged (missing), and the remedy it now gives (put the declaration on a line of its own) is correct for that shape as well.

End-to-end, on this card's own thread

Running the reader over the real comment thread of #16931: before, the row carried no quotation at all; after, it quotes the triage bullet that spells the key mid-line. The state is absent in both readings, unchanged.

Ablation

Fix committed first. Mutating the detector's single decision (return true for a key preceded by non-decoration) to return false, proved on disk by blob hash 60b7562c… → 71d6a50b…:

• self-test exit 1, 10 of 319 cases red — every case that pins the new diagnostic.
• The controls in that batch stayed green by design: state still missing, no value, tally still one not-read card, the standing refusal still printed. Those hold in both worlds, which is what makes them controls.
• Restored with git checkout HEAD -- PATH; blob back to 60b7562c…, git diff HEAD empty, git status --porcelain empty, self-test exit 0 / 319 pass.

Test evidence

• pnpm check:pm-clause2-carriers — exit 0, 319 cases (was 299 on origin/main; 20 added across the two existing batteries, which carry floors rather than equalities, so SELF_TEST_BATTERIES is untouched).
• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 35 commands; all 35 run; reconciled with --ran: 35 derived famil(ies) accounted for — 35 run, 0 NOT-MEASURED, exit 0.
• Seven of the 35 first returned exit 3 — PREREQUISITE NOT MET (pnpm install had not run in this worktree). That is NOT MEASURED, not a finding; all seven were re-run after install and returned 0.
• Also run, because they name this file: pnpm check:pm-half-states (0), pnpm check:pm-widening-tells (0). Their reference is to this file's decoration tolerance as a precedent; that tolerance is unchanged.
• No separate vitest suite exists for this script — its suite IS its --self-test, wired as check:pm-clause2-carriers.
• eslint . --no-inline-config --format json at 59adc6d0: 6438 files, 0 errors, 0 warnings. Whole population, not a narrowing.

Changeset

None. scripts/pm/** ships in no package's files[] and the repo root package is private, so nothing published moves. lint.yml states the case in its own words: "this PR edits a CI-internal script" is the textbook skip-changeset case.

验收备注 — noted, not filed

• ⛔ No collision with [finding] check-clause2-carriers T2 fires on a replaced string property value as "a new member of a closed set", and the C5 row it raises cannot be cleared except by declaring Clause-② yes on a change that does not widen #16943. That card is the C5 widening-tell logic (T2 on a replaced string value) in a different region of this file. This diff touches the near-miss reporter, readClause2Line, cardDeclaration's return object, the missing branch of c2DeclarationUnreadable, and two self-test batteries — no overlap, and SELF_TEST_BATTERIES is not edited. ⛔ [finding] check-clause2-carriers T2 fires on a replaced string property value as "a new member of a closed set", and the C5 row it raises cannot be cleared except by declaring Clause-② yes on a change that does not widen #16943 is not addressed here and remains open.
• The claim comment on this card is not a claim comment to the predicate. It spells the key as backticked `Claim:`, and CLAIM_COMMENT_MARKER is a line BEGINNING Claim: (blockquote tolerated, backticks and bold not). So [finding] a Clause-②: key that is not at the start of its own line is invisible to BOTH of check-clause2-carriers' regexes — read as missing, with a remedy that names only spelling #16931's thread reads absent rather than missing. Measured, not inferred. That strictness is deliberate and documented in this file, so this is a note about the carrier, not a defect in the reader — the seat that wrote the claim is the one who can move it. ⛔ Not filed, ⛔ not edited by this seat, and ⛔ the declaration was not filled in on anyone's behalf.
• The absent row's own remedy sentence is deliberately left alone: what that thread owes is the claim comment, which is correct whatever the near miss looked like. It gains only the quotation it never had.

🤖 Generated with Claude Code

https://claude.ai/code/session_012GKcPZbMoGq7WPzKLfRBTU

---

Generated by Claude Code