- @xenova/transformers 2.17.2 in a Web Worker; Opus-MT en↔es ONNX q8 packs
self-hosted on the relay VPS (/models + CORS); ONNX Runtime 1.14 WASM
self-hosted in public/ort/ (37MB) loaded from ../ort/.
- Hard network guard: only model GET/HEAD (no query) and /ort/ assets allowed.
Message text travels only via postMessage; 9 tests prove fetch is never called
with the text and every other request is blocked.
- TranslateMessage in DMs + crew chat; Settings auto-translate off by default;
memory-only translations; per-model download progress; 12 UI strings in all
24 locales.
- Both model directions verified from the hosted files; tsc, 138 tests, build
and smoke green.
- tinyld (MIT, lazy, local) + script rules + a weighted
Catalan/Galician/Basque/Spanish/Portuguese rule for the close/short cases
tinyld confuses. No network, no logs, no storage; conversation/contact prior
supported.
- Lingua with 24 languages produced a 99 MB WASM, so it was rejected as a
detector; tinyld + the gl rule covers all 24 at a fraction of the size.
- 28/28 language samples pass, including a test that fetch is never called.
- Phase 2: on-device Opus-MT (English pivot) in a worker, memory-only
translations, model packs on Blossom, sentinel/network-guard test.
- Full 8,703-key Hindi dictionary in src/i18n/hi (7 namespace batches),
placeholders verified, only 2% identical strings (brand/protocol names).
- Register hi in Language/LANGUAGE_LOADERS, detectLanguage (hi-IN),
LanguageToggle (🇮🇳, हिन्दी) and dateFnsI18n (hi-IN).
- Lazy-load Noto Sans Devanagari for hi via lib/localeFonts.ts (renamed from
cjkFonts.ts) and add it to the CSS font stack; other locales don't download it.
- check-i18n covers hi: 24 dictionaries / 25 codes at 100%; tsc/tests/build green.
