docs(spec): correct the published 17.4.0 entry that restated the retired M9.2 promise

[os-bill]

Part of #17849

Clause-②: no

The second half of #17849, split out of PR #18557 because AGENTS.md's Documentation Guardrails row for packages/*/CHANGELOG.md requires it: an already-published entry is amended in a dedicated docs-only PR, ⛔ never as a rider on code changes. This diff is exactly one file and adds no source, no schema and no export.

#17323's ruling item 2 orders it: 「packages/spec/CHANGELOG.md: an erratum line under the entry that promised M9.2 (the #17026 shape — correct the published text, note the date), ⛔ not a rewrite of history.」

What was corrected, and where

One entry: ef3a138 (feat(spec)!: an evaluated expression slot requires a non-blank source) under the already-published ## 17.4.0. Line numbers re-derived on this tree at 79a046f8c, ⛔ not carried over from the card or from the earlier round:

line	published text	disposition
1096	the blockquote quoting EVALUATED_EXPRESSION_SOURCE_REQUIRED, parenthetical (the canonical persisted form of phase M9.1)	left exactly as shipped — it is a faithful quote of what 17.4.0 published. A paragraph under it says what the constant reads now.
1103	「its docblock declares that ast becomes required in build output at phase M9.2」	corrected in place, old words kept as a marked quotation
1119	「has no evaluable form under M9.1」	corrected in place, old words kept as a marked quotation

One adjacent clause falsified by the same ruling is corrected in the same stroke and named here rather than smuggled: the bullet also read 「when AST-only evaluation lands」, which presupposes the retired promise. It now reads 「if AST-only evaluation is ever chartered」 — the wording the two pending changesets already carry. Leaving it would have left a 「when it lands」 sitting beside 「no promise of becoming required」 in one bullet.

One dated erratum line closes the entry, carrying the in-repo tail this repository already uses in five places (packages/spec/CHANGELOG.md ×3 at :2457, :2882, :5277, packages/lint/CHANGELOG.md:1214, packages/metadata-protocol/CHANGELOG.md:134):

*Erratum, 2026-09-17 — the M9.1 / M9.2 phase promise this entry restated was retired by the
ruling on #17323 (2026-09-12) … (Corrected after publication, #17849.)*

⛔ No new entry at the top, ⛔ no version heading added (git diff -U0 | grep -c '^+## ' → 0), ⛔ nothing this release published is changed.

⚠️ The card's grep criterion cannot be met, and the ruling is why

The card sets git grep -l 'M9\.[12]' origin/main -- packages → 0. After this PR that file still carries four hits, and every one of them is required by the ruling's own 「⛔ not a rewrite of history」:

:1096  the as-shipped blockquote                       (deliberately untouched)
:1109  "…becomes required in build output at phase M9.2"   inside `As published, that sentence continued "…"`
:1130  "no evaluable form under M9.1"                      inside `As published that clause read "…"`
:1142  the erratum line itself, naming the retired promise

Not one of them is a live assertion of the promise — they are the quotation marks the #17026 shape puts around it, plus the erratum that retires it. A zero would require deleting the published words, which is precisely the rewrite the ruling forbids. ⇒ the criterion and the ruling are not jointly satisfiable, and the ruling governs. Reported rather than forced.

⚠️ skip-changeset — checked against #18375 before relying on it

The label is applied. Before applying it I re-read the refusal #18375 is about, scripts/check-empty-changeset.mjs, and it does not reach this PR:

• that gate has two rules and both take the .changeset/ diff and nothing else — rule 1 fires on a newly ADDED empty-frontmatter changeset, rule 2 (scanForeign()) on a MODIFIED or DELETED changeset that exists on the merge base;
• the DELIBERATE-CORRECTION text 「no label and no diff shape makes that safe」 is scoped, in its own words, to 「the note you rewrote describes behaviour THIS PR changed」 — a pending release note, i.e. a .changeset/*.md;
• this diff contains no .changeset/ path at all (git diff --name-only origin/main...HEAD → one line, packages/spec/CHANGELOG.md). So the label suppresses no refusal that could have fired here, and the finding's hazard is absent rather than accepted.

What the label IS doing is the documented job: Check Changeset requires an added changeset from every PR, with no path filter, so skip-changeset is the only instrument for a diff that releases nothing of its own — the same instrument the #17026 ruling named (item 3) and PR #17896 spent for the same shape.

Verification

Exit codes landed to disk before reading, ⛔ never through a pipe.

run	verdict
derived gate families for this one path	55 derived · 51 run green · 4 NOT MEASURED
--ran reconciliation with per-family exit codes	55 derived famil(ies) accounted for — 51 run, 4 NOT-MEASURED (4 DERIVED from a recorded exit 3)
check-release-section-coverage (plain · --self-test · --strict)	exit 0 ×3 — 7 published minors across 2 GA majors, every one still covered
pnpm check:release-notes, pnpm check:release-page-status	exit 0
pnpm lint (eslint . --no-inline-config, whole repo)	exit 0
pnpm check:nul-bytes + a direct control-character sweep of the file	exit 0 / no hits

The 4 NOT MEASURED are check:dts-closure, check:dual-build-cjs-loads, check:lean-entry-closure and check:sourcemap-no-sources-content — all exit 3 = PREREQUISITE NOT MET (they read the dist/ of packages this tree never built). ⛔ Neither a pass nor a failure; CI's Build Core runs them.

No test is owed and that is measured, not assumed. Four test files name a CHANGELOG.md path and every one of them excludes it: compliance-families-retirement.test.ts skips CHANGELOG.md, both action-owner-key-single-source.test.ts files list it under covers as 「a published CHANGELOG is the record of the removal itself」, and template-consistency.test.ts passes :(exclude)**/CHANGELOG.md. No test reads the bytes this PR moves.

The changeset version survival question is already answered and is not re-run here. #17026's round measured it with a lit control: a manual edit inside an already-compiled section SURVIVES changeset version (marker present before and after, while the tool provably re-wrote both files — a new version section prepended, the package version advanced, 250 pending changesets consumed). That reading stands; re-running it would buy nothing.

Acceptance notes

Confirming reading for the finding recorded on #17849: pnpm check:cross-package-test-inputs is green (exit 0) on this tree, which has no packages/spec/dist because a markdown-only diff needs no build. Same gate, same commit base, opposite verdict from the built tree — a third independent leg for the same cause. ⛔ Not filed again; it is already #18353 and #18440.

Noted, not filed: nothing else. Successor for the M9.9b / M9.5 / M9.7 ROADMAP citations elsewhere in the tree: none — no PR or person is routed to those files by this work, and the ruling names only the M9.1 / M9.2 phase promise.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

⚠️ 1 changed file(s) yielded no anchor (packages/spec/CHANGELOG.md), so the pages documenting them are NOT COVERED by this run — this is not a clean bill of health for those files. Nothing else in this diff resolved to a documentable surface (no symbol, route or SDK anchor derived from 1 changed package(s)).

What this run could not see

• 1 changed file(s) yielded no anchor (packages/spec/CHANGELOG.md) — pages documenting those are invisible to this run
• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.

Coarse fallback — 136 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json 79a046f8cdf085d95200826ee9bb2fa6584bc3d5 → packageMentionDocs.