chore: release v0.29.4

[mmcky]

v0.29.4 — patch. Promotes [Unreleased] for three merged PRs that change what the engine does:

PR	What it ships	Affects
#291	Sync and the rebase replay carry the target's translated _toc.yml part captions forward, matching parts by the lectures they share, and change only the caption lines (#254)	Every edition with translated captions; one lecture-adding sync reverted 14 of them on lecture-python.zh-cn
#289	A new lecture the run failed to deliver is removed from the sync PR's _toc.yml, by splicing out its lines, and the PR body lists it under Files Failed to Translate (#156)	Lecture-adding syncs where a new lecture is refused or fails to fetch: the target build no longer breaks on a dangling toctree entry
#335	ml rule 19: "have already seen / met" takes the completive -കഴിഞ്ഞു	Malayalam drafts and syncs

Also on main since v0.29.3, with no change to the engine: Malayalam experiment arms and tooling (#331, #333) and the .dev/ → .qe/dev/ move (#314).

Version bumped in package.json / package-lock.json, CHANGELOG promoted under a fresh [Unreleased] header, and a .qe/dev/STATE.md entry. The rest of .qe/dev/STATE.md is brought up to date with the #289/#291 merges. No source change; dist-action/ rebuilds byte-identical (engineVersion is a runtime read of package.json). npm test 1635/1635; lint, format:check and check-dev-refs pass.

What the gate can and cannot see. The harness's _toc.yml fixtures are flat (no parts:, no captions). Scenarios 17, 18 and 20 therefore exercise #291's new target-TOC fetch, which now fails on any status but 404, and its no-parts no-op. They do not exercise the caption carry-forward, which unit tests cover, including byte-identity on a QuantEcon-style TOC and a test of #291 and #289 together. #289's filter fires only if a new lecture actually fails, as scenario 17 did on the .ml lane at v0.29.1.

Not in this release: #346's French and Malayalam items (#344, #347, #327, #324, #325) stay with v0.30.0; Japanese (#69) waits on #337 and #338, then ships with #340.

After merge: tag v0.29.4 → §4a E2E gate on the tag and the 4b rate check → move v0.29 and v0 → @v0 smoke → GitHub release. Tallies will be posted here.

🤖 Generated with Claude Code

[Copilot]

🔵 Needs a closer look

Cutting this release gates floating-tag moves (@v0/@v0.29) and billed downstream runs and is explicitly a maintainer decision, and the STATE.md verified: freshness marker should be bumped before it ships.

1 open finding

• STATE.md verified date and resume heading are stale · New

What changed in this PR

This is a mechanical patch release PR (chore: release v0.29.4) that promotes three already-merged, engine-affecting PRs plus accumulated non-engine changes into a published version. It fits the repository's documented Release Checklist (AGENTS.md): bump the version, promote [Unreleased] in the changelog under a fresh empty header, and update the internal project-state notes.

Changes:

• Bump version 0.29.3 → 0.29.4 in package.json and package-lock.json.
• Promote the [Unreleased] changelog section to ## [0.29.4] - 2026-10-09 (carrying #291/#254 TOC caption preservation, #289/#156 failed-file TOC entry removal, #335 ml rule-19 completive, and ml experiment arms), leaving a fresh empty [Unreleased] header.
• Update .qe/dev/STATE.md: add a v0.29.4 In-flight entry, move #289/#291 to "Recently landed", and adjust the resume checklist.

File	Description
package.json	Version bump to 0.29.4
package-lock.json	Matching self-version bump to 0.29.4 (both occurrences)
CHANGELOG.md	Promotes [Unreleased] to [0.29.4] - 2026-10-09 under a fresh empty header
.qe/​dev/​STATE.md	Records the v0.29.4 cut and moves #289/#291 to "Recently landed"

Notes from verification: version bumps are consistent across both manifests; the changelog promotion keeps an empty [Unreleased] header and the three engine-changing PRs are all represented; the 2026-10-09 date equals the current date (not a future date). The one finding is that STATE.md's verified: freshness marker (first line) and the "Resume here (2026-09-23 …)" heading were not bumped to reflect this refresh, which is a documented convention in .qe/dev/README.md.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[mmcky]

§4a gate on the v0.29.4 tag (fcee846): passed. 4b rate check: over threshold on zh-cn, judged pre-existing. Released.

	zh-cn	fa	ml
sync runs succeeded	28 / 28	28 / 28	28 / 28
translation PRs delivered	28	28	28
verdicts at engineVersion: 0.29.4	28 / 28	28 / 28	28 / 28

Census 9/9 workflows at @v0.29.4; 28 source PRs; 84/84 sync runs, none failed. Two runs per lane read as skipped, not failed: the reset closing the previous gate's test-translation-sync#976 fires a pull_request and an issue_comment event that the workflows skip. Scenario 17 (new-document-toc) delivered game-theory.md on all three lanes (test-translation-sync.zh-cn#1009, .fa#394, .ml#300). As the PR body says, the harness TOC fixtures are flat, so this gate exercised #291's target-TOC fetch and its no-parts no-op, not the caption carry-forward.

§4b rate check (rate-check.sh --ref v0.29.4): game-theory.md refused 0/12 on fa and ml, 2/12 on zh-cn, over the threshold of one. Both zh-cn refusals carry the same structural-parity signature: directives 3 → 2, with {code-cell} and {math} transposed. That is #203's signature exactly. Whether this was a regression:

• The zh-cn request is byte-identical at both tags. translator.ts and glossary/ are unchanged, the zh-cn LANGUAGE_CONFIGS entry is unchanged, and the init path imports none of the modules that changed (index.ts, pr-creator.ts, sync-orchestrator.ts, toc-captions.ts).
• Control draws, zh-cn, game-theory.md:

Engine	Draws today	2026-09-23	Total
v0.29.3	0 / 24	0 / 12	0 / 36
v0.29.4	2 / 12, then 1 / 24		3 / 36

The difference is not significant (Fisher two-sided p ≈ 0.24), and with identical request bytes it is sampling variance. Pooled, the latent rate is about 4% (3 / 72). At that rate, the "more than one refusal in 12" rule fails about 8% of the time by chance. So @v0 was already carrying this defect at v0.29.3, and holding the tags back would not have protected zh-cn from it. In the action path, #289 now makes a refusal like this cost one lecture rather than the target build. Measurement posted on #203.

Checklist 5: v0.29 and v0 moved 16e50f6 → fcee846. @v0 smoke (scenario 01): 3/3 sync runs succeeded; verdicts report engineVersion: 0.29.4, engineRef: v0 on all three lanes (test-translation-sync.zh-cn#1020, .fa#404, .ml#309). Checklist 6: GitHub release published, title = the tag.