Fix TTS voice being ignored/inconsistent in read-aloud and video export

[davior]

Summary

Fixes the reported bug: setting the read-aloud voice to "Rex" (a fal.ai voice) did not take effect in AI Assistant read-aloud or in MP4 export, and some rendered videos changed voice partway through.

Three separate root causes:

• AI Assistant read-aloud / voice mode: AIConversationPanel.tsx and useVoiceMode.ts both called useTextToSpeech() with no model option. When the fal.ai fallback path is used, the frontend defaults an unset voice to the hardcoded DEFAULT_TTS_VOICE ("Aria") instead of the user's configured voice. Fixed by passing the settings store's current voice through, matching how EditorView.tsx and SpeechSettings.tsx already do it.
• Article-to-video export: the render worker's _tts_caller never passed the user's selected voice into synthesize_tts_bytes at all, so the fal.ai path always fell back to the current TTS model's first listed voice — never the user's actual choice. Added load_selected_voice() to read the per-user voice setting and pass it through for every render.
• Voice changing mid-video: when the read-aloud provider is "auto" and both a Deepgram and fal.ai key are configured, provider selection is re-resolved on every synthesis call, and a transient Deepgram failure silently falls back to fal.ai for just that one call. For a narration job that makes one call per text chunk, this could swap a single chunk to a different engine/voice mid-render. The provider is now resolved once per render job and pinned for the whole render, so a mid-job hiccup now fails the render (which is safe to retry) instead of producing a video with an inconsistent voice.

Test plan

• Verified the fal.ai voice-resolution path (_fal_tts, synthesize_tts_bytes) and confirmed the new voice/provider_override plumbing is used consistently by the render worker.
• Reviewed the frontend hooks (useTextToSpeech, useVoiceMode, AIConversationPanel) to confirm the selected voice is now threaded through everywhere useTextToSpeech is instantiated.
• Manual verification in a running instance (set voice to a non-default fal.ai voice, confirm it's used in AI Assistant read-aloud and in an exported MP4) was not possible in this sandbox — dependencies (fal_client, pydantic, frontend node_modules) aren't installed, so this couldn't be exercised end-to-end. Please verify manually before merging.

🤖 Generated with Claude Code

https://claude.ai/code/session_01R54PjuaaN6QnvxVCcxiuVs

---

Generated by Claude Code