ml: stop the exercise rule priming invented directives — v0.29.1's gate failed scenario 17 (latent since v0.28.0)

[mmcky]

v0.29.1's §4a gate came back 83 of 84. The miss is scenario 17 (new-document-toc) on the .ml lane, and a re-run against the same tagged bytes failed identically — so the floating tags were not moved and no GitHub release was published. v0 and v0.29 still point at v0.29.0 (a6fda54). This PR is the repair; it ships as v0.29.2.

What failed

The fixture game-theory.md writes its exercises as a plain ## Exercises heading and a numbered list, with no directives. The model wrapped them in {exercise-start} … {exercise-end}. The structural-parity guard did its job and refused the file (directives 3 → 5), which left the sync PR carrying a _toc.yml entry for a lecture it had not delivered — the #222 / #289 class. The primer is the ml exercise-verbatim rule: it names every exercise directive at length and never says they must already be in the source.

It is a latent defect, not only a regression

Two draws were not enough to tell, and nor were six per arm (they read 1/6 against 5/6 and looked like a clean regression). Twelve draws per arm on the fixture, refusals by the parity guard:

Rules	Glossary	Refused
v0.29.0	v0.6.0	5 / 12
v0.29.0	v0.7.0	8 / 12
v0.29.1	v0.6.0	11 / 12
v0.29.1	v0.7.0	11 / 12
v0.29.1 + this PR	v0.7.0	0 / 24 (two wordings, twelve each)

The defect has been there since the rule arrived in v0.28.0: the v0.28.0, v0.28.1 and v0.29.0 gates each passed scenario 17 on a single draw with roughly a 40% chance of failing. The #315 rules aggravate it to about 90%. The crossing places that in the rules rather than the glossary; why longer comma and plural rules should have this effect is not known, and the log says so rather than guessing.

The change

One sentence added to the exercise rule — still 28 rules. It says the rule concerns directives the source already contains, that no directive, fence or label may be added, and that a plain ## Exercises heading and list keeps its English text and its source structure. A first wording told the model to translate that list; it never did at either version (0 of 19 written draws), and the editor's ruling covers all exercise content, so the wording was changed to ask for the behaviour the ruling wants. Reverting #315 was not considered: those are the editor's answers, and the defect predates them.

Harness

Local dist-action bundle through the action path, on the fixture PR that failed (test-translation-sync#907): delivered 2 of 2 — test-translation-sync.ml#221, #222, each with game-theory.md, its state file and _toc.yml. The tagged v0.29.1 failed 2 of 2 on the same PR.

Worth carrying forward

A single-draw gate cannot see a 40% defect; it passed three releases. Retries or N draws on the new-document scenarios belong with #260 / W1 (#259). And a rule that lists markup should say "already in the source".

Suite 1590 green (+2); lint, format:check, check-dev-refs clean; dist-action/ rebuilt.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟡 Changes recommended

Release-status documentation is internally inconsistent (STATE/CHANGELOG) and should be clarified before shipping the patch release notes.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 2

Open (2)

• Changelog mixes tagged-only and released wording · New
• Release header implies v0.29.1 was published · New

What changed in this PR

This PR updates the Malayalam (ml) language rules to explicitly scope the “exercise-verbatim” rule to directives already present in the source, preventing the model from inventing {exercise-*} directives around plain ## Exercises sections (fixing the scenario-17 .ml gate failure that blocked the v0.29.1 release).

Changes:

• Extend the ml exercise-verbatim rule to forbid introducing any new exercise directives/fences/labels when the source uses a plain ## Exercises heading + list.
• Add a targeted Jest assertion to lock in the new scoping language while keeping the rule count at 28.
• Rebuild dist-action/ and document the incident/fix in CHANGELOG.md and .dev/ notes.

File	Description
src/​language-config.ts	Adds the scoping sentence to the Malayalam exercise-verbatim rule to prevent directive invention.
src/​__tests__/​language-config.test.ts	Adds regression coverage asserting the rule explicitly forbids adding directives and keeps rule count unchanged.
dist-action/​index.js	Updates the bundled Action output to reflect the source rule change.
CHANGELOG.md	Documents the fix under [Unreleased].
.dev/​STATE.md	Records the v0.29.1 gate failure context and that the repair ships as v0.29.2.
.dev/​log/​2026-09-21-v0291-gate-scenario17.md	Adds a detailed incident log entry for the gate failure and resolution.

---

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