feat: add Malayalam (ml) language support — keep-English-dominant policy

[mmcky]

Summary

Adds Malayalam (ml) as a target language, following the native-speaker review on #70 by Adisankar Manoj Thanuja (Kerala finance educator). Refs #70 — this is the first increment; it does not close the tracking issue.

Unlike zh-cn/fa, which fully translate technical terms, Malayalam uses a keep-English-dominant policy. Kerala has no Malayalam-medium STEM/finance education infrastructure and a globally-mobile, English-fluent workforce; learners use English technical terminology natively, so translating or transliterating it reads archaic ("government gazette," not classroom). Technical terms stay in English with Malayalam grammar wrapping around them; only everyday connective words are translated.

Policy

Category	Treatment
Technical terms — economics, finance, statistics, programming	keep English
Acronyms (GDP, RBI), institutions (Federal Reserve)	keep English
Proper names (economists, researchers)	keep English/Latin
Everyday non-technical words (country, year, before)	translate to Malayalam
Transliteration	effectively unused

The characteristic pattern is Malayalam grammar attached to English roots inline — economy-യിലെ, bond-ന്റെ, asset classes-ൽ — and English verbs with a Malayalam auxiliary (process ചെയ്ത്, return ചെയ്യുന്നു). This English-root-plus-Malayalam-inflection behaviour is an LLM strength a sentence-level NMT system can't replicate.

What's in this PR

• src/language-config.ts — ml entry with six keep-English rules (keep technical terms English, attach Malayalam morphology to English roots, keep proper names in Latin, enforce per-document consistency, classroom register with sparing parenthetical-first-use).
• glossary/ml.json — 60-term seed: 47 technical terms pinned to English (ml == en) + 13 everyday words translated, every Malayalam value taken directly from the reviewer's worked examples (nothing machine-guessed).
• dist-action/ — rebuilt bundle (npm run build), including dist-action/glossary/ml.json.

Status: 🟡 draft / seed pending native-speaker validation

This is intentionally a seed for a calibration batch: run real lectures on this config (Opus, ml) and have Adisankar flag each over-translation so we tune the glossary/rules from real output rather than theory. The primary risk for ml is consistency (a term must be handled identically on every occurrence), not over-translation.

Decisions still open (see #70)

• Proper-names policy — defaulted to keep-English/Latin; reviewer to confirm.
• Parenthetical-on-first-use convention (ബന്ധം (relationship)) — keep sparing.
• v1 stays zero-schema-change (ml == en); the per-term treatment field is deferred unless calibration shows we need it.

Follow-ups (not in this PR)

• Per-language Opus default for ml + add claude-opus-4-8 to the model allowlist in src/inputs.ts.
• Tests for the ml config/glossary.
• Docs: docs/user/language-config.md, docs/user/glossary.md, glossary/README.md.

Verification

• npm run build — clean; dist-action/ bundle regenerated and committed (CI freshness guard).
• npm run lint — 0 errors (pre-existing warnings only).
• npx jest on language-config, inputs, sync-orchestrator, translator, integration, cli-smoke — 166 tests pass.
• glossary/ml.json — valid JSON, 60 terms (47 keep-English, 13 translated).

🤖 Generated with Claude Code