feat: replace local Supertonic TTS with Voxtral

[AnthonyRonning]

Summary

• replace local Supertonic 3 inference with encrypted OpenSecret speech synthesis using voxtral-tts and neutral_female
• preserve progressive desktop playback, fully prebuffered iOS playback, cancellation, and existing platform audio-session behavior
• remove Supertonic models, dependencies, native commands, download and preference UI, and obsolete documentation
• safely remove exact legacy tts_models directories at launch while retaining ONNX Runtime for PDF OCR
• add focused coverage for text sanitization and chunking, synthesis errors and cancellation, playback state, and legacy-model cleanup

Dependencies

• OpenSecret backend: Proxy paid Voxtral speech through Tinfoil opensecret#241
• OpenSecret SDK: feat: add encrypted Voxtral speech synthesis APIs OpenSecret-SDK#90

Maple remains pinned to @opensecret/react 3.2.1. This PR is intentionally being opened before the SDK release and version update. Current frontend and web-build CI fail at TTSContext.tsx because the published SDK does not yet export synthesizeSpeech; a follow-up must update frontend/package.json and frontend/bun.lock after the SDK release. The backend must also be deployed before remote synthesis works outside the local managed workspace.

Related workspace troubleshooting documentation: https://github.com/OpenSecretCloud/opensecret-workspaces/pull/10

Validation

• 257 frontend tests passed
• production frontend build and TypeScript checking passed using the path-linked in-flight SDK
• ESLint passed with 0 errors and the existing 12 warnings
• 8 native legacy-cleanup tests passed
• full native suite: 206 passed / 1 model-backed OCR test ignored
• repository-native Rust lint, formatting, and diff checks passed
• managed-stack smoke:
  • Free account rejected with HTTP 403
  • Pro account returned a valid 272,684-byte, 5.68-second WAV
  • desktop control completed Read aloud → Stop speaking → Read aloud
  • a legacy-model fixture was removed on launch

Known gaps

• Physical iOS/TestFlight locked-device playback and Windows/Linux runtime playback have not been tested.
• The iOS cross-build attempt stopped in upstream native dependencies because the local Nix/Xcode toolchain mixed macOS and iOS flags, before compiling Maple code.
• A debug macOS app was produced and smoke-tested; updater signing could not finish without the private signing key.
• Human listening remains the final voice-quality check.
• Web and Android TTS remain intentionally unsupported.
• A stricter nonstandard all-target Clippy invocation reports two lints in unchanged agent permission tests; the repository-native just rust-lint passes.