v0.27.0

The second Malayalam native-review round is encoded, and every translator call now uses prompt caching.

Malayalam round 2 (#296, #297)

The editor of record reviewed functions on lecture-python-programming.ml#7 — 118 suggestion blocks on 118 of 123 Malayalam prose lines, all applied and merged. The round-1 rules held (no recurrence of below-deixis, ordinal, light-verb, -യുടെ or suffixed-English flags), so this release carries the new classes the round surfaced:

• Rules 18 → 24: terminal punctuation on every Malayalam paragraph (a colon before the cell, list or equation it introduces); sentence-initial capitalisation of retained-English words; the hortative teacher's voice (നമുക്ക് … -ആം, never നമ്മൾ … -ും); consistent accusative -നെ and no additive -ഉം on list items; fixed renderings ("a given" → തന്നിരിക്കുന്ന, "consider" → നോക്കാം, "useful" stays English); the sandwich-sentence rule; three more English-retained connectives; and the maintainer's 2026-09-01 ruling that exercise statements with probabilistic or mathematical content stay English, hybrid sentences included (D-2026-09-01-ml-exercise-statements-stay-english).
• Glossary v0.4.0 (61 → 77 terms): the ordinary words the editor kept moving back to English, the line/lines pin, and four everyday words he moved to Malayalam.
• First deterministic ml lints in ml_metrics.py (LINT, never FAIL): terminal punctuation, lowercase-initial sentences, banned renderings, a future-hortative watch — calibrated seed 24/19/16/6 → the editor's reviewed text 1/1/2/1.
• Regeneration-checked: a fresh claude-sonnet-5 pass lands 17 of 26 signatures at the reviewed form; the two mechanical classes (punctuation, bullet capitalisation) under-deliver from prompt alone and are the case for a deterministic post-processing step (#260).

Prompt caching (#292, #293)

Translator prompts are split into a stable, cache_control-marked block (rules + glossary) and a volatile block (per-file instructions + content); repeat calls in a run read the 10–12K-token prefix at the cached rate. tokensUsed and input-tokens now exclude cached tokens — total prompt size is input + cache-creation + cache-read.

Also in this release

• tool-review-injection/: certify the review gate by injected-defect catch rate (#285) — first run 72% M0-core catch rate.
• runReview extracted from src/index.ts and tested (#278, #169 slice 1).
• E2E harness targets carry .translate/config.yml with the production editors: block again (#275).

Full detail in CHANGELOG.md.