Let another app read a whole video in one pinned voice

[davior]

Gecko Video Creator's renderer (M2) reads its narration through Notes' speech, one chunk per request, so voices, keys and the TTS cache stay in one place. It needs two things the endpoint couldn't give it:

• The account's voice when none is chosen. TTSRequest.model defaults to a fal.ai voice, which isn't the same thing.
• The engine and voice pinned after the first chunk. Notes' own render worker already does this: a Deepgram hiccup partway through must not read one chunk on fal.ai in a different voice.

What changes

• New optional fields. POST /api/settings/speech/tts takes provider (auto | deepgram | fal) and voice, where null means the account's voice.
• Shared resolution. With provider set, both are resolved by video/worker.resolve_render_speech, which is pulled out of _tts_caller so the worker and the endpoint pin the same way.
• Echoed back. The response names the resolved pair in X-Speech-Provider and X-Speech-Voice, for GVC to send back on every later chunk.
• Old requests unchanged. Without provider, a request means exactly what it did: read-aloud, export and the frontend are untouched.
• Spec doc updated. docs/gecko-video-creator.md §9.1 marks this row done.

Tests

• Three new tests in tests/test_video_voice.py, with providers faked as the existing tests do:
  • "auto" gives Deepgram in the account's voice, with the headers set;
  • a pinned engine and voice are kept;
  • a request without provider is unchanged, with no headers.
• The worker's existing pinning tests still pass through the extracted helper.
• pytest -q: 1351 passed, 1 skipped.

🤖 Generated with Claude Code

https://claude.ai/code/session_016fiRJAt34ag62iPiuMkNhm

---

Generated by Claude Code