ml: encode the editor's round-3 answers — rules 27 → 28, glossary v0.7.0 (two shipped guesses corrected)

[mmcky]

Encodes Adisankar's answers to the round-3 questions (lecture-python-programming.ml#22, answered 2026-09-19). Small payload: ml rules 27 → 28, glossary v0.6.0 → v0.7.0, one decision record, two lints and one mechanical repair. Two of his answers overturn a guess — one of which v0.29.0 had already shipped.

What he said, and what changes

Item	His answer	Change here
Comma between -ഉം items	Always — phrases and single words, "for reading clarity"	The clause-boundary rule, part (c), said phrases only: corrected, with single-word examples. The suffix rule's own example (a green border-ഉം a blinking cursor-ഉം) contradicted it: corrected too
draw	വരയ്ക്കാം is natural Malayalam and worked better in that lecture	Nothing encoded, on purpose. He also changed വരയ്ക്കുന്ന to draw ചെയ്യുക in the same lecture, so both forms are his. A test pins that draw is not in the light-verb list
Plural of a retained noun	English plural + Malayalam suffix; agrees -ുകൾ after a hyphen renders badly	Its own short rule (objects, function calls-ൽ, never object-ുകൾ) rather than another clause in the longest rule of the set, plus a lint
prefer	Keep in English	Glossary pin, light-verb form
provide	Both acceptable; നൽകുന്നു read better	Glossary provide → നൽകുക, superseding the v0.4.0 pin; the context note says the other form is never an error in reviewed text
Which ordinary words stay English	"Would a Kerala student say this word in everyday conversation?" is "a fair test"	Not in this PR — see below
Further Reading	Intentional: reference material pointing outside the lecture	D-2026-09-21-ml-further-reading-boundary-is-the-reference-section; the prompt rule's wording already matched. Moving it into code is now unblocked under #260

Measured, not assumed

Three draws of numpy on this branch beside the three archived v0.29.0 draws — same source commit, model and command; numpy is held out (no rule quotes it). Arm: experiments/ml-benchmark/arms/2026-09-21-round3-answers-numpy/.

	v0.29.0 (3 draws)	this branch (3 draws)
provide ചെയ്യ…	5 / 5 / 6	0 / 0 / 0
hyphenated -ഉം pairs carrying the comma	0 of 4	2 of 6
Malayalam plural on an English noun	0	1 (functionality-കൾ — a mass noun)

The glossary change is deterministic. The corrected comma rule lands about a third of the time, missing the same two lines in both draws that had them — the terminal-punctuation finding again: a rule he states as "always" does not come reliably from the prompt. So ml_repair.py (the #260 prototype) gains the comma as its third mechanical repair. The one plural is on a noun with no natural English plural, the same case as logic-ുകളെ on the reviewed functions page; that is a question to him (asked on lecture-python-programming.ml#24), not yet a rule.

Harness

Local dist-action bundle on the .ml lane: test-translation-sync.ml#191 from fixture PR test-translation-sync#890. Delivered; 87-term glossary loaded; the plural rule fired on the fixture's own concepts-കൾ → concepts; the hortative held. It also missed the comma on its one -ഉം pair and opened a sentence with lowercase modern — the second harness run in a row to do that. Both are lintable misses of rules already present, which is the case for the lint-and-repair step (#260, #301) over more prompt.

Deliberately not here: the rule-2 rewrite

He accepts the everyday-speech test, which unblocks the rewrite the round-3 report proposed (the principle in the prompt, the word lists in the glossary). It is held for a held-out run of the blind pairwise judge: rule 2 is the rule behind the ordinary-word pendulum of rounds 2 and 3, the confirmation was mild and to a leading question, and draw — which we had listed under "stays English" — shows the test is not sharp.

Round 4 is regenerated after this releases

Decided 2026-09-21, reversing the first call of the day. lecture-python-programming.ml#23 (numpy, v0.29.0 draw 3) has had no human activity, and the arm above changed the arithmetic: branch draw 1 is cleaner than the draft he was sent on every lint (bare endings 0 vs 3, comma-less -ഉം pairs 0 vs 2, provide ചെയ്യ… 0 vs 6, banned 2 vs 3). His review time is the scarce resource, so he should read the engine's best current output. He has been asked to hold off, on ml#22 and on ml#23. No branch draw goes to him — every round has been generated at a published @v0 — so the sequence is the round-3 one: merge this → v0.29.1 with its §4a gate → three draws at @v0, best by lint, ml_repair.py applied and disclosed → force-update ml#23 and retitle it. If he says he has already started, ml#23 stays as it is and the known flags (the comma-less pairs, any provide flip) are excluded from the convergence residue.

Checks

Suite 1588 green; lint, format:check and check-dev-refs clean; dist/ and dist-action/ rebuilt. A patch release is enough — nothing here is a feature.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟢 Approval recommended

The changes are consistent, tested, and scoped to rule/glossary encoding plus lint/repair tooling, with only minor wording nits noted.

Review effort: Lite
Findings: 2

Open (2)

• Fix mid-sentence capitalization in glossary text · New
• Update hard-coded rule-count range in test name · New

What changed in this PR

This PR updates the Malayalam (ml) localization configuration to encode the editor’s round-3 answers (issue lecture-python-programming.ml#22), including a ruleset bump (27 → 28) and a glossary bump (v0.6.0 → v0.7.0). It also extends the benchmark tooling with new lints and a deterministic repair to address a comma rule that remains unreliable when enforced purely via prompting.

Changes:

• Update ml prompt rules to reflect the editor’s clarified comma usage for -ഉം … -ഉം coordination and add a new short plural rule for retained Latin-script nouns.
• Update the Malayalam glossary to v0.7.0 (notably provide → നൽകുക, and pin prefer in English).
• Extend ml_metrics.py with two new lints and ml_repair.py with a third deterministic repair for missing comma insertion between adjacent hyphen-suffixed -ഉം items.

File	Description
src/​language-config.ts	Encodes editor answers: comma rule strengthened; adds new plural rule; updates commentary for ml#22 decisions.
src/​__tests__/​language-config.test.ts	Updates glossary/rules expectations and adds assertions covering the new ml#22 encodings.
glossary/​ml.json	Bumps glossary to v0.7.0 and updates term pins (e.g., provide → നൽകുക, add prefer).
experiments/​ml-benchmark/​scripts/​ml_repair.py	Adds deterministic comma insertion repair for adjacent -ഉം pairs without commas.
experiments/​ml-benchmark/​scripts/​ml_metrics.py	Adds lints for Malayalam plural suffix on Latin-script nouns and missing commas between adjacent -ഉം pairs.
experiments/​ml-benchmark/​arms/​2026-09-21-round3-answers-numpy/​README.md	Documents measured impact of the rule/glossary changes on held-out numpy draws.
experiments/​ml-benchmark/​arms/​2026-09-21-round3-answers-numpy/​numpy-ml22-draw1.md	Adds benchmark artifact for draw 1.
experiments/​ml-benchmark/​arms/​2026-09-21-round3-answers-numpy/​numpy-ml22-draw3.md	Adds benchmark artifact for draw 3.
dist-action/​index.js	Rebuilt action bundle reflecting updated language config and glossary behavior.
CHANGELOG.md	Adds [Unreleased] entries describing the ml rules/glossary bump and new lint/repair behavior.
.dev/​STATE.md	Updates in-flight status for the ml#22 encoding work.
.dev/​log/​2026-09-21-ml-round3-answers.md	Adds a dev log entry capturing the editor answers, decisions, and measurement summary.
.dev/​decisions/​D-2026-09-21-ml-further-reading-boundary-is-the-reference-section.md	Records the clarified boundary decision for “Further Reading” scope.
.dev/​decisions/​D-2026-09-18-ml-further-reading-lists-stay-english.md	Adds a pointer note to the new boundary decision record (without editing the original decision text).

---

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