iOS voice input transcribes non-English speech with the English model

[nichtlegacy]

Problem

On iPhone, voice input transcribes non-English speech very poorly. Speaking German produces mostly English-sounding nonsense.

The cause is the language the app hands to Apple's SpeechTranscriber. voiceTranscription.ios.ts uses Intl.DateTimeFormat().resolvedOptions().locale, which Hermes reads from [NSLocale currentLocale]. That is the app locale, not the user's language. T3 Code only ships an English localization, so on a German iPhone the app locale is en-DE, SpeechTranscriber.supportedLocale(equivalentTo:) resolves it to an English model, and German speech runs through the English model. This affects every non-English user, and there is no setting to override it.

Quick way to see it on the current App Store build: dictate an English sentence, then a sentence in your device language. English comes out fine; the other does not.

Proposed fix

#13312 (small, JS-only, no native build needed):

• Pick the transcription language from the user's preferred languages (AppleLanguages, the same list as Locale.preferredLanguages) and fall back through the list when SpeechTranscriber does not support one. The app locale stays as the final fallback.
• Apply automatic word spacing on insertion to every space-delimited language, not just en. Otherwise fixing the locale would make German users lose the spacing they currently only get because of the bug.

Apple's recommended API for this on iOS 26 is Locale.preferredLocales, matched through SpeechTranscriber.supportedLocale(equivalentTo:). Doing that natively would mean extending the existing @react-native-ai/apple patch and shipping a native build. The PR reads the equivalent list through React Native's built-in Settings module so it can ship as an update. Happy to go either way.

The PR is unit-tested but not yet verified on a physical device.

Possible follow-ups (not in the PR)

• A voice input language setting (Automatic / explicit language) for people whose device language differs from the language they dictate in.
• An optional OpenAI-compatible transcription endpoint (custom base URL + API key, e.g. whisper-1, gpt-4o-transcribe, or a self-hosted Whisper server), ideally environment-backed so the key stays on the server and web/desktop/Android could use it too. docs/internals/voice-input.md already anticipates environment-backed transcription.