feat: add Russian localization and independent STT languages

docs/localization.md

@@ -0,0 +1,54 @@
+# Language support
+
+LearnHarness supports English and Russian. Settings → General contains independent
+interface and speech recognition language selectors. The microphone also has a
+speech language selector. Each offers System default, English, and Русский.
+
+System default uses the preferred browser language (the operating system's
+preferred language in the desktop app); Russian locales select Russian and other
+locales select English. Explicit choices persist in `settings.json` as
+`general.language` and `general.sttLanguage`. Removing a field restores automatic
+selection. Changing one preference preserves the other general settings.
+
+The interface changes immediately, including accessibility labels and the desktop
+menu. New AI teaching, questions, plans, practice, graph descriptions, research,
+figures, and demos use the interface language. Existing conversations and imported
+source material retain their authored text. A session retains the language of its
+latest foreground turn so automatic follow-ups can use the same language after a
+restart; an explicit app language preference takes precedence.
+
+Speech recognition uses Deepgram Nova-3 with an explicit `language=en` or
+`language=ru`. A recording captures the selected language when it starts and keeps
+that language through reconnections. The selector is disabled while recording;
+changes elsewhere take effect on the next recording. Audio capture, deferred
+transcript insertion, stopping, and cancellation retain their existing behavior.
+
+## Maintaining translations
+
+The pure localization functions live in `src/core/i18n.ts` and are exported through
+the core barrel. English and Russian catalogs are in `src/core/locales/en.json` and
+`ru.json`. English source phrases are message keys. Add every new app-owned message
+to both catalogs, use named interpolation parameters, and use plural entries with
+`Intl.PluralRules` for counts. Russian entries may use `one`, `few`, `many`, and
+`other` forms. Keep code, provider/model names, identifiers, and user-authored text
+out of translation lookups.
+
+React components use `useI18n()` from `components/i18n.tsx`. The root provider gets
+server-resolved initial settings and locale to preserve hydration. General-setting
+writes share a serialized read-modify-write path; other surfaces use the existing
+settings-saved signal. Server and agent code receive an explicit language rather
+than depending on a mutable global translation state.
+
+Some existing protocol errors arrive preformatted in English. Localization can
+recognize cataloged application templates at the display boundary, preserving the
+canonical error values used for classification and any third-party diagnostics.
+Raw technical instructions, code, and user overrides remain editable source data.
+
+The desktop shell has its own small catalog in `electron/language.cjs`, packaged
+with the shell. Its narrow `lh:language` channel accepts only `en` and `ru` from a
+trusted application window. Russian prose uses Noto Sans because Source Sans 3
+has no Cyrillic subset. Explicit font choices continue to take precedence.
+
+`tests/i18n.test.tsx` covers preference validation, automatic selection, independent
+STT requests, catalog parity and interpolation, Russian plural forms, SSR locale
+isolation, translated dialogs and exports, agent instructions, and native labels.