Partial update of a saved voice config

OpenAPI-visible TTS overrides accepted by voice-config CRUD.

The linked voice_id remains the authority for which provider member is valid. The service resolves that row and validates this payload again against the provider-specific discriminated union, including model support and the Inworld TTS-2-only expressive-speech constraint.

Silero VAD tuning. Always applied for Deepgram Nova; ignored otherwise.

Smart-turn-v3 tuning. Always applied for Deepgram Nova; ignored otherwise.

Config for the AI's opening turn (no user input yet).

Lives inside ChannelConfig — first-speaking is per-channel.

Mode-agnostic fields:

• first_speaker: "agent" opens the call with the rendered message; "user" skips the greeting (caller speaks first). Non-voice channels treat the inbound user message as the "user" opener and ignore this field at runtime.

Greeting source (only applies when first_speaker="agent"):

• static mode: text is required and emitted verbatim (templated via the shared {{var}} engine, src.shared.templating.render).
• dynamic mode: instructions are sent to the model router with the agent persona to render an opening line (instructions are templated). When omitted/empty, instructions falls back to DEFAULT_DYNAMIC_FIRST_SPEAKING_INSTRUCTIONS so operators who just want a generic friendly opener don't have to author copy.

Voice-only fields (ignored on non-audio channels):

• interruptible: whether the caller can barge in mid-greeting.
• ai_disclosure_text: regulatory disclosure (e.g. CA SB 1001), played as an uninterruptible TTS frame BEFORE the greeting. Disclosure plays even when first_speaker="user" because the regulation requires the disclosure regardless of who speaks first.