[finding] Four more copies of the falsified `paths-ignore` premise survive #3857's three named sites (changeset-presence.yml, check-changeset-no-major.mjs and its test)

[yinlianghui]

Out-of-scope finding, recorded while implementing #3857 / #3859 (PR pending). Filed unassigned, per the report-don't-fix discipline: #3857 names exactly three sites and its dispatch scope is those three, so these four were left untouched rather than swept in.

Fact

#3857 established that ci.yml and lint.yml no longer carry paths-ignore on their pull_request trigger — objectui#3523 step 2 moved the path decision into the jobs (Decide whether this change needs a full run), leaving paths-ignore only on push. It then named three texts that still assert the old shape. A keyword sweep for paths-ignore across the tree at origin/main @ 2fea4d2fa finds four more carrying the same falsified premise, none of them named on that card:

Location	What it still says
.github/workflows/changeset-presence.yml:11-15	"ci.yml and lint.yml both list .changeset/** under paths-ignore, so a PR that adds ONLY a changeset starts nothing else" — a near-verbatim copy of the scripts/check-changeset-presence.mjs paragraph #3857 does name
scripts/check-changeset-no-major.mjs:22	"nothing executed it, and no workflow even looked at .changeset/** (both ci.yml and lint.yml list it under paths-ignore)" — the sentence is historical, but the parenthetical is present tense
scripts/__tests__/check-changeset-no-major.test.ts:19-22	"ci.yml and lint.yml both paths-ignore .changeset/**, so a changeset-only PR started no workflow at all"
scripts/__tests__/check-changeset-no-major.test.ts:120-123	"ci.yml cannot host this check: a changeset-only PR matches its paths-ignore twice over ('**/*.md' and .changeset/**), so no job in it would ever run"

The two test-file occurrences are comments only. The assertions below them are sound and stay green — expect(ci).toContain("paths-ignore:") passes because the push trigger genuinely keeps its copy — so nothing here is a broken gate; it is the explanation attached to a working gate.

Why it is worth a card rather than a shrug

This is the same class #3857's own triage graded, in its words: agent-facing operational assertions that are live-disproven, "the 'decide by it, decide wrong' class, with the cost already sampled once (PR #3856's author wrote a false CI expectation from AGENTS.md and got 16 checks)". That card was promoted from finding to pm:queue on that reading.

There is one extra cost specific to these four, which did not apply when all seven copies said the same wrong thing: once #3857's PR lands, the repository will state both shapes. A reader who lands on changeset-presence.yml's header first has no signal that the corrected text three files away supersedes it — the two read as equally authoritative, and the stale one is the one physically closest to the gate it explains.

The right fix is the same one #3857 specifies, applied to these four: state that paths-ignore remains only on push, that the pull_request path decision lives in the Decide whether this change needs a full run step, and restate each gate's rationale as the reason that holds today (both changeset gates are correct as built — the conclusion outlives its premise, because the in-job exclusion list still covers markdown and .changeset/**, so those workflows start, report, and skip every expensive step on exactly the PR these gates exist to judge).

Live counter-proof, re-measured today

• PR docs(skills): console-development.md 顶部注记的搬家归因改成 04-21→04-23 的 commit 链,cccdf84d7 降为其中一步 (#3737) #3856 — one .md file, 16 checks (the original sample).
• PR docs(agents): state commit-then-revert as the required order for reverse verification (#4301) #4339 — one line added to AGENTS.md, merged 17:40Z on 2026-08-11: 17 check runs on its head commit, read from the API, including Lint, Type Check, Test (shard 1..4/4), Internal Docs Link Check, Skill Guide Path Check, Control Byte Scan and Changeset Declaration.

How to re-check

grep -rn 'paths-ignore' .github/workflows scripts AGENTS.md | grep -v node_modules
sed -n '1,20p' .github/workflows/ci.yml          # paths-ignore on push only; :12 says why
sed -n '29,42p' .github/workflows/lint.yml       # same, :37

Dedupe

Searched open issues in this repo for paths-ignore, changeset-presence, check-changeset-no-major and changeset-guard: the only hit is #3857 itself, whose three named sites are disjoint from the four above. No twin.

Relation

• [finding] 三处 agent-facing 文本仍称「ci.yml/lint.yml 把 **/*.md 与 .changeset/** 列进 paths-ignore,只加 changeset 的 PR 不启动任何 workflow」—— #3523 step 2 已把 pull_request 的 paths-ignore 删掉,实测 md-only PR 起了 16 个 check #3857 — parent in substance, not in scope: same premise, four sites it did not enumerate. Its PR corrects only the three it names.
• docs(skills): console-development.md 的「Retired names」表称 config/metadataTypeRegistry.ts「从未存在过」—— git 可证它确实以该路径落过地(d601ef693 加入 / ff9a0d9e9 删除) #3859 — the other half of the same docs-truth batch; different file, different fact.

---

Generated by Claude Code

[yinlianghui]

CLAIM — PM seat session_017Qqyix2QcnpUC9XeYVDzx3, branch claude/issue-4369-paths-ignore-residue, one dev agent, one PR, Fixes #4369. finding → pm:queue, promoted on the card's own reading: same live-disproven class #3857's triage graded, with the new both-shapes cost now real since PR #4371 landed (merged 22:4xZ) — the stale copy in changeset-presence.yml is the one physically closest to the gate it explains.

Dispatch shape: mirror #4371's landed corrections at the four sites the card enumerates (workflow header, script docblock parenthetical, two test-file comment blocks). Comment-only, zero behavior change, assertions untouched. Surface is disjoint from all five in-flight tasks.

---

Generated by Claude Code