feat(scripts): count the changeset bodies check-changeset-claims cannot reach

[baozhoutao]

Part of #9140 — this delivers ONE of the card's two defects and deliberately leaves the other open. It does not close the card.

Which defect this delivers

#9140 separates two things, and they have different fixes:

defect	delivered here
(a) blind spot — a WENT-FALSE shape the gate cannot see	partly — made measurable, ⛔ not closed. See "why not closed".
(b) delivery — the gate fired four times and changed nobody's behaviour	⛔ no — every option is a policy choice. See "left open".

Reproduction first — the card's measurement, re-run on this branch's base

Base 9ba7e9c3a2. Each falsifying merge re-run on its own parent..merge range, node scripts/check-changeset-claims.mjs --base MERGE-CARET --head MERGE (MERGE-CARET spelled in words: the caret is the parent operator):

falsifying merge	pending changeset it falsified	gate named it
c5fbe0b99	.changeset/5293-view-sort-order-spelling.md	yes
adbda1bed	.changeset/7165-grid-dependent-values.md	⛔ no
d3499b315	.changeset/7655-chatbot-registration-authoring-faces.md	yes
ad66d79a1	.changeset/7727-conditional-formatting-record-scope.md	yes
0601af1ad	.changeset/8934-chatter-feed-affordance-only.md	yes

⇒ the card's finding reproduces exactly: 4 of 5 named, the 5th invisible. Every run exited 0.

⚠️ One correction to the card. Its second table's absolute span counts do not reproduce under the gate's own namedFiles() — I read 1 / 0 / 4 / 4 / 2 where the card reads 1 / 10 / 16 / 3 / 0. Different instrument (the card counted spans "containing a slash or ending in a source extension", undeduped; namedFiles() requires both and dedupes). The finding is unaffected and reproduces: the one the gate could not see is the one — uniquely — with zero.

Why the coordinate was NOT widened

I measured the widening rather than assuming it. The cheapest in-idiom candidate stays inside the gate's existing machinery: treat a backticked span equal to the stem of exactly one tracked file as naming that file (ObjectGrid → packages/plugin-grid/src/ObjectGrid.tsx). Same instrument, same unambiguous-resolution filter, no symbol table.

Re-measured over the 25 most recent first-parent commits on my base, paragraphs per commit:

coordinate	median	mean	max
file named unambiguously (current)	1	3.8	25
+ stem, symbol-only bodies (candidate)	0	5.3	34
union	1	8.6	59
(header's already-rejected runner-up)	8	12.6	56

The first row reproduces the header's own row (median 1 / mean 4.2 / max 23) — my instrument agrees with theirs, which is what makes the other rows comparable.

⛔ The union's tail exceeds the runner-up the header rejected. And the crux is worse than the aggregate suggests:

• .changeset/7165-grid-dependent-values.md enters through exactly one spelling: ObjectGrid.
• At that merge's base, ObjectGrid is backticked by 41 of 759 pending changesets.

⇒ seeing that one true positive means asking the seat to re-read up to 41 paragraphs on every commit that edits the grid. That is the release-time wall in a smaller costume — the exact shape the header rejected on a measurement — on a channel that half one of #9140 just measured at zero for four. Widening is a readability-budget decision, it is genuinely open, and ⛔ I did not take it. The header records the trade; the numbers stay on the card.

What this PR does instead

--audit now counts the population the diff-mode check can never reach for any diff: bodies that publish a CHANGELOG paragraph but spell no name the gate can resolve — either naming no file at all (symbol-coordinated, the 5th's shape) or naming only files that resolve to none or to many. --audit --blind lists them.

The card established that population by counting backticked spans by hand, once, on one card. It is now re-derived on every run — AGENTS.md rule #9 applied to the gate's own blind spot.

The gate's coordinate, its findings and its report-only nature are unchanged; --audit's existing lines are unchanged.

Left open — for whoever revisits this

• (b) delivery. Nothing here improves it. Making the gate block is fenced by the card and by the header's own measurement; routing its output somewhere a seat must acknowledge (a PR comment, a review request) is a policy choice about a channel measured at 0-for-4. ⛔ Not invented here — finding(gate): check-changeset-claims fired on 4 of the 5 WENT-FALSE claims objectui#9065 had to repair by hand, and nobody acted on any of them — the 5th it could not see at all #9140 has no triage ruling, and this needs one.
• (a) closing the blind spot. Three live options, now with costs: change nothing; widen to the stem coordinate at the volume tabled above; or add a compact second report tier that lists changesets rather than quoting paragraphs. The header records that the choice is open.

Verification

• scripts/__tests__/check-changeset-claims.test.ts — 28 passed (3 new).
• Firing negative control: a symbol-only body is counted in the blind population; the same body with one resolvable file name added leaves it (silent drops by exactly 1, with both runs pinned to the same publishing population so neither side can be reading nothing).
• The blind spot is pinned as a limit: the diff-mode gate stays silent about a symbol-only body even when the change edits the file that body is about — with a control in the same run showing the gate was not simply silent about everything.
• Ablation (committed first, mutation proved on disk by blob hash, restored byte-identically): deleting the silent counter turns 2 of the 3 new tests red; restore verified by an empty git diff HEAD and a matching blob hash.
• check-changeset-claims / -no-major / -fixed / -overwrite / check-control-bytes / check-changeset-presence: all exit 0.
• eslint on the 2 changed files: 0 errors, 0 warnings (narrowed from the 4935 files eslint's own config selects; no type-aware linting is configured, so this diff cannot move the verdict on any untouched file). tsc -p tsconfig.scripts.json: exit 0, and --listFiles confirms the new test file is in that program.
• Changeset declares no bump (empty frontmatter): tooling and tests only, nothing is released.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FhBNJcLRZLe8M87VcUgpKr

---

Generated by Claude Code