Split on-device STT into hardware-aware providers

[ComputelessComputer]

Separate Soniqo and Apple Speech from Anarlog Cloud, migrate existing selections, and drive recommendations and QA coverage from shared hardware metadata.

---

Note

Medium Risk
Touches core transcription provider selection, config migration, and batch/live routing for Soniqo and Apple Speech; mis-routing could break recording or downloads for existing users.

Overview
On-device speech-to-text is no longer folded into a single “Anarlog local” notion: Soniqo and Apple Speech are first-class STT providers (separate from Anarlog Cloud), with existing current_stt_provider / model selections expected to migrate to the right provider IDs.

The desktop app uses isOnDeviceSttModel (replacing isAnarlogLocalSttModel) for paths like local download/status polling in NotificationProvider and STT health, so only true on-device provider+model pairs trigger local STT behavior.

Settings → AI → STT gains hardware-aware picking: models carry recommended_memory_bytes, recommendOnDeviceModel picks a default from available RAM, and the UI can label the pick with “Best for this Mac” (new Lingui string on select.tsx).

QA critical UX now requires running the on-device matrix for every STT model shown on the QA Mac (not one representative), recording model ID, hardware recommendation status, and per-model pass/fail in release checklists.

^{Reviewed by Cursor Bugbot for commit 2a95832. Bugbot is set up for automated code reviews on this repo. Configure here.}

[netlify]

✅ Deploy Preview for old-char canceled.

Name	Link
🔨 Latest commit	2a95832
🔍 Latest deploy log	https://app.netlify.com/projects/old-char/deploys/6a6ef29bd39ec100083952f2