feat(ml): encode native review round 2 — glossary v0.4.0, rules 18 → 24, first deterministic ml lints

[mmcky]

Encodes the second inline native-review round — 118 suggestion blocks on QuantEcon/lecture-python-programming.ml#7 (functions), all applied and merged there as c1200fa, every one dispositioned in QuantEcon/project-translation reports/2026-09-01-ml-functions-review-disposition.md. Tracker: #296. The round-1 rules held (no recurrence of below-deixis, ordinal, light-verb, -യുടെ or suffixed-English flags), so this PR carries the new classes the round surfaced.

Rules (src/language-config.ts, ml 18 → 24): terminal punctuation on every Malayalam paragraph (a colon before the cell, list or equation it introduces); sentence-initial capitalisation of a retained-English word (Functions-നെ, Built-in, Return statement); the hortative teacher's voice (നമുക്ക് … -ആം, never the plain future നമ്മൾ … -ും); consistent accusative -നെ on light-verb objects, no additive -ഉം on list items, `a`, `b` എന്നിവ; the fixed renderings he corrected on every occurrence ("a given" → തന്നിരിക്കുന്ന, "consider" → നോക്കാം, spelling കുറച്ചുകൂടി, "useful" stays English); the sandwich-sentence rule (a claim interrupted by code cells is stated once, before the first cell); three more English-retained connectives ("In particular", "Obviously", "Basically"); worked examples for the pointer sentences the model missed three times; and the maintainer's 2026-09-01 ruling that exercise statements with probabilistic or mathematical content stay English, hybrid sentences included — decision record .dev/decisions/D-2026-09-01-ml-exercise-statements-stay-english.md, with the pure-programming negative example named so the rule does not over-apply.

Glossary (glossary/ml.json v0.4.0, 61 → 77): the ordinary words the editor kept moving back to English (useful ×4 — never ഉപയോഗപ്രദമായ — provide, exist, automatically, create, improve, full, stage, answer, successive), the line/lines pin the #273 regeneration flagged (14× variance regression), and four everyday words he moved the other way (straightforward → എളുപ്പം, contrived → കൃത്രിമം, limited → പരിമിതി, facilitate → സഹായിക്കുക).

Lints (experiments/ml-benchmark/scripts/ml_metrics.py): the round's four deterministic classes as a LINT bucket (never FAIL until #189 Phase 3 decides what graduates) — terminal punctuation, lowercase-Latin-initial sentences, a banned-renderings list drawn from the rules, a future-hortative watch. String checks, so unaffected by #240 and #229. Calibrated on real text: the pre-review seed scores 23/17/15/6, the editor's reviewed text 1/0/2/1, and every residual is a form he himself kept.

Regeneration check (same recipe as #273; archived under experiments/ml-benchmark/arms/2026-09-01-round2-rules-sonnet5/): a fresh claude-sonnet-5 pass of functions from source 81d9734 lands 17 of 26 signatures at or past the reviewed form — every fixed pair, the list-item and sandwich forms, the connectives, all four exercise-scope retentions and both pointer sentences. The two purely mechanical classes under-deliver from prompt alone (bare endings 23 → 16, bullet capitalisation 17 → 5): the strongest case yet for a deterministic post-processing step (bare Malayalam ending before a {code-cell} → :; lowercase Latin word at the head of a prose line or list item → capitalised), W2-shaped (#260) and left out of this PR deliberately; the lints detect both on every PR meanwhile. Logged one instance each: set transliterated (a real pinned-term FAIL the gate caught), refer calqued.

Held for the editor's answers on QuantEcon/lecture-python-programming.ml#12 (deliberately not encoded): "For example," as a discourse rule; glossary pins for already, name, example(s), work.

ml-scoped only — no other language's rules move; the Stage 4 shadow-window freeze (cross-language rubric, floors) is untouched. Full suite 1538 green; lint and format clean; dist-action rebuilt. Docs (docs/user/glossary.md, docs/user/language-config.md) and CHANGELOG updated. Release after merge so the rules reach production before round 3 (matplotlib) is generated.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Encodes the second Malayalam native-review round into the engine’s ml rules/glossary, and adds deterministic (non-blocking) lint signals to the Malayalam benchmark tooling so future sync PRs can surface these classes consistently.

Changes:

• Extend ml.additionalRules (18 → 24) to cover the new review-round classes (punctuation, casing, hortative voice, fixed renderings, sandwich-sentence, exercise-scope extension).
• Bump Malayalam glossary to v0.4.0 (61 → 77 terms), including new keep-English pins and a few additional everyday-word translations.
• Add “round-2” lint reporting (LINT-only) to ml_metrics.py, plus tests/docs/changelog updates and a new decision record.

Reviewed changes

Copilot reviewed 11 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/language-config.ts	Adds round-2 Malayalam rule text to ml.additionalRules.
src/tests/language-config.test.ts	Extends tests to assert v0.4.0 glossary additions and presence/absence of key round-2 rule snippets.
glossary/ml.json	Bumps Malayalam glossary to v0.4.0 with new keep-English pins and added everyday translations.
experiments/ml-benchmark/scripts/ml_metrics.py	Introduces “round-2 lints” (terminal punctuation, lowercase-initial, banned renderings, hortative watch) as LINT output.
experiments/ml-benchmark/arms/2026-09-01-round2-rules-sonnet5/README.md	Archives the round-2 regeneration arm summary and signature table.
experiments/ml-benchmark/arms/2026-09-01-round2-rules-sonnet5/functions.md	Stores the regenerated Malayalam functions output used for the arm’s scoring/comparison.
docs/user/language-config.md	Updates Malayalam documentation to describe the second inline review round and the new rule classes.
docs/user/glossary.md	Updates built-in glossary table to reflect Malayalam v0.4.0 and new term count.
dist-action/index.js	Rebuilds the committed Action bundle to match updated source rules.
CHANGELOG.md	Adds an Unreleased entry describing the Malayalam round-2 encoding and lints.
.dev/decisions/D-2026-09-01-ml-exercise-statements-stay-english.md	Records the maintainer ruling extending the English-retention scope to math/probability exercise statements.

---

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