The H2 half-state predicate is blind to two MORE claim spellings in live use — measured, 9 of 22 pm-tracked cards are false positives after #7488's fix #7553

Filed by the dev on #7488 / #7440 (PR #7551) from the fleet measurement that card required. Unassigned — recording, not claiming. No domain:* label: same two artifacts as #7488, routing is the triage seat's call.

What #7488 fixed, and what it deliberately did not

#7488 ruled direction 1: h2AssigneeNoClaimComment now accepts an optional markdown blockquote marker, so the shape SKILL.md step 4 tells every seat to write (> Claim: …) reads as a claim. That card ⛔ explicitly barred widening the regex any further — "do not widen the regex so far that ordinary prose containing the word claim starts matching (its current strictness is a feature)" — so what follows was measured and left alone, on purpose.

The measurement

Sample: all 18 open pm:dispatched cards (the exact population the H2 sweep reads) plus the 4 recently-closed cards #7488's table names. 22 cards, every comment thread read through the API on 2026-08-11.

So after #7488 lands, 9 of these 22 cards still report as H2 half-states while carrying a perfectly good claim comment. That is a higher false-positive rate than the 1-in-8 #7488 was filed about.

⚠️ A correction to #7488's own body, for the record. It says of the other seven H2 hits: "the other seven are true findings — they carry no claim comment in any form." Its table is literally correct (it measured exactly two spellings, and neither is present), but the generalization is wrong: all seven carry claim comments, in the two spellings above. #7488's acceptance criterion is unaffected — none of the seven uses a blockquote, so all seven are still reported after its fix, which is what PR #7551 verifies.

Why this is a ruling, not a dev-level tooling call

Unlike the blockquote form, neither of these spellings is what SKILL.md tells seats to write, so "the predicate cannot read the documented shape" — #7488's whole argument — does not apply. There are two coherent directions and they point opposite ways:

1. Converge the fleet on the documented shape. The predicate stays strict; the seats writing **Claim** — and 认领 are the thing that is wrong, and the fix is protocol discipline (plus, perhaps, an H-invariant that reports a claim-shaped comment the predicate cannot parse, so the divergence is loud instead of silent). Costs: every already-posted claim in those spellings stays unreadable, and 5 of the 9 are from the two most active PM seats.
2. Teach the predicate all four spellings. Costs exactly what finding: the H2 half-state predicate cannot read the claim block SKILL.md tells every seat to write #7488 warned about — each spelling added is regex width, and 认领 plus bold-without-colon is a long way from "a line that begins with Claim:". The strictness that makes a green H2 mean something starts eroding.

Note the interaction with the maintainer's 2026-08-08 English-on-GitHub ruling: the four Chinese-spelling cards are from 2026-08-05..09, i.e. mostly older than or contemporaneous with that ruling, so direction 1 may already be happening on its own. Worth measuring the trend by date before choosing.

⛔ Not fixed in PR #7551 — that PR is scope-barred to the blockquote marker by #7488's own instruction.

Refs: #7488 (the blockquote half), PR #7551 (the fix), #7412 (neighbour — the same script's live sweep cannot run in a PM/dev session container; independently re-confirmed during this work: api.github.com/repos/** answers 403 with and without the container token).