test(ci): fixture the no-major changeset guard and wire it into the exemption-free self-test step (#6923)

[os-project-manager]

Fixes #6923

scripts/check-changeset-no-major.mjs was the one member of the changeset-gate family with no --self-test, which is why PR #6917 could not wire it into lint.yml alongside the other two. This adds one, following the idiom of check-empty-changeset.mjs and check-adr-0087-registration.mjs (pure exported judge, fixtures that drive the real decision, missing-input-is-a-failure), then uses those fixtures to measure the enforcing half instead of reasoning about it.

Premise re-verified on origin/main first

Both halves of the card still held at d3e53f2d8:

• grep -n "self-test\|selfTest" scripts/check-changeset-no-major.mjs — no output.
• .changeset/pre.json is "mode": "pre", so the real scan takes the RC exemption branch and exits 0 on every run.

Which half the fixtures exercise, and what is still unexecuted

Stating this plainly because it is the thing most easily overstated.

• Exercised by fixtures: every decision the script makes, both halves. The refactor extracts a pure judge({ changesets, pre }) and a pure render(verdict), so the enforce verdict — the offenders report and process.exit(1) — is now produced and asserted on every PR, message and exit code alike.
• Still unexecuted on CI: the enforcing half. .changeset/pre.json is unchanged and still "mode": "pre" (this seat performs no release actions, release workflow: publish pushes tags + npm but its version commit never reaches main — twice now (rc.3 c6a52d3, rc.4 a10cbc77); landing the commit must be part of the publish lane #6170), so every real CI invocation still exits 0 through the exemption branch. Fixtured is not executed. What changed is only that the branch is now driven by fixtures, in a job with no PR-level label exemption. The script's header says this in the same words so the next reader does not read one as the other.

The measurement that changed the diff

The dialect fixtures are measured against @changesets/parse@0.4.3 — the version this repo resolves — rather than written from the regex. That found a real divergence, so this PR is not fixtures-only:

input	@changesets/parse	guard, before	after
"@objectstack/spec": major	major	caught	caught
'@objectstack/spec': major	major	caught	caught
docs: major (unquoted)	major	caught	caught
CRLF	major	caught	caught
a leading blank line before the fence	major	MISSED	caught
"@objectstack/spec": MAJOR	throws	caught	caught
no closing fence	throws	caught	caught
"@objectstack/spec": major # keep	major	MISSED	MISSED (#7004)

The parser required its fence on line 1 (if (lines[0]?.trim() !== '---') return []), so a changeset opening with one blank line declared nothing to this guard while changesets honoured its major and would have promoted the entire lockstep group. Both sibling parsers already skipped leading blanks, and all three carry a comment saying the three read the same block — so this was also the one place that comment was false. One line, aligned with the siblings; it can only ever add offenders, and today it adds only ::notice lines.

The trailing-comment row is a family-wide gap in a shared regex with a different consequence in each of the three scripts, so it is filed (#7004) rather than fixed here. The fixture pins the current behaviour with a message telling the next author to flip it, not delete it.

Reverse verification — predictions written before running

Ten ablations, each with its expected failure count stated up front. Nine matched exactly; one did not, and is reported as it happened rather than reshaped.

#	ablation	predicted	observed
R1	delete the leading-blank-line tolerance	3	3 — the three leading-blank fixtures
R2	pre?.mode === 'pre' becomes !== 'exit'	5	crash, not 5 — see below
R2b	same, with the null-deref repaired	5	5 — the exact five named
R3	enforcing half exits 0	1	1
R4a	unwire from check:changeset-gate-self-tests	1	1
R4b	same, seen by the sibling's pin	1	1 (check-empty-changeset)
R4c	wire the REAL scan into the shared step	2	2
R5	delete the real scan from pr-automation.yml	1	1
R6	stop excluding .changeset/README.md	2	2
R7	rename lint: so the slice anchor goes stale	2	2
R8	unreadable dir returns an empty Map, not null	1	1

R2 diverged and the divergence is worth recording. The prediction was five assertion failures. What happened was a TypeError: Cannot read properties of null (reading 'tag') — the loosened test lets pre === null reach the exempt branch, which then dereferences pre.tag. Still red, but by crashing before any assertion ran, so it proved less than intended. R2b repeats it with pre?.tag and produces exactly the predicted five: no pre.json at all, pre.json that did not parse, pre.json with no mode key, mode: some future spelling, mode: PRE (wrong case) — with mode: exit correctly staying green.

R7 is the vacuous-green trap this family keeps hitting, checked deliberately after reading check-required-contexts.mjs (#6983). The wiring block slices lint.yml by anchor; a stale anchor yields an empty slice that every assertion below would pass against. R7 renames the job, and the guard fires by name (the lint: job could not be sliced out of lint.yml — its anchors went stale) rather than passing silently.

Wiring

check:changeset-gate-self-tests gains a third entry. Two halves land in deliberately different places:

• the self-test joins lint.yml's ESLint job, which has no PR-level exemption — what 观察单：changeset 家族三个 check-*.mjs 的 --self-test 只在可被 skip-changeset 整体豁免的 job 里跑，改这些脚本的 PR 恰好豁免自己的自检 #6509 / PR fix(ci): run the changeset family's --self-test in lint.yml, out of reach of skip-changeset (#6509) #6917 built the step for;
• the real scan stays in pr-automation.yml, because its allow-major and skip-changeset exemptions are deliberate. Moving it into lint.yml would silently revoke the escape hatch this script's own error message prescribes. Asserted in both directions.

check-empty-changeset.mjs is edited for one reason: its consumer assertion stated in its own message that this script is "deliberately absent: it has no --self-test to run". That was true when #6917 wrote it and false as of this commit, so leaving it would make the repo assert something false. Replaced with an assertion that this script IS wired, and its --self-test-on-every-invocation regex extended to cover the new entry (R4b/R4c confirm both are load-bearing).

What was measured and deliberately NOT acted on

Per the card's boundary — fixtures may document the enforcing half; changing release-mode semantics is not this PR's to do. .changeset/pre.json is untouched.

The enforcing half reads the whole .changeset stock with no branch point, unlike its diff-scoped sibling. Measured at d3e53f2d8: 1552 changesets on disk, 171 of them declaring major (222 package entries), still present because pre-mode changeset version records consumed changesets in pre.json (1279) rather than deleting them. changeset pre exit rewrites the mode to "exit" (@changesets/pre@2.0.2), so from that commit until the final Version PR merges, every unlabelled PR would fail this guard listing 171 files it never touched, with allow-major as the only route out. That is #6129's direction by another route and it decides what the guard means, so it is filed (#7005) for triage, not improvised here.

Findings filed (unassigned, out of this PR's scope)

• A trailing YAML comment hides a bump entry from all three changeset frontmatter parsers #7004 — a trailing YAML comment hides a bump entry from all three changeset parsers. False GREEN in this script, false RED in check-empty-changeset (a valid changeset reads as empty-frontmatter). Measured against the real parsers by extracting their source, not by re-implementing them.
• check-changeset-no-major.mjs judges the whole .changeset stock, not the PR's diff — at changeset pre exit it reds every unrelated PR (171 files measured) #7005 — the stock-scoped enforcement described above.
• [finding] An unreadable .changeset/ makes check-changeset-no-major.mjs print a tick and exit 0 (#4690 shape) #7006 — finding: an unreadable .changeset/ still exits 0 (check:react-declaration-parity 是唯一没接进任何 workflow 的源码审计门禁,且无 MANIFEST 时静默 skip 退出 0 —— 它现在永远不可能红 #4690 shape). Not changed here, but now distinguishable (no-changeset-dir vs clean), which is the prerequisite for changing it.

Verification

$ pnpm check:changeset-gate-self-tests
✓ check-empty-changeset --self-test: 65 assertions over real temp git repos (real scan() path)
✓ check-adr-0087-registration --self-test: 107 assertions over real temp git repos (real scan()/assertInputs() path)
✓ check-changeset-no-major --self-test: 70 assertions (frontmatter dialects measured against
  @changesets/parse + the pre/exit exemption switch in both directions + the #4690 reader pins + the wiring).

$ node scripts/check-changeset-no-major.mjs      # real scan, output unchanged
✓ Changesets is in pre-release mode (tag: rc) — `major` bumps are the expected product of an RC window; ...

$ pnpm lint          # exit 0
$ pnpm check:nul-bytes
check-nul-bytes: OK (scanned 6473 tracked text file(s); ... no raw ASCII control bytes).

No changeset: this PR touches scripts/ and the root package.json only and publishes nothing, so it carries skip-changeset.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 9, 2026 8:39am