feat(desktop): OpenAI-endpoint transcription + configurable STT model in Settings

[agentforce314]

Two follow-ups to the voice fixes, per request.

1. Use the real OpenAI endpoint with an OpenAI key

STT now defaults to the provider's canonical audio endpoint (openai → api.openai.com/v1, groq → api.groq.com/openai/v1) instead of the provider's chat base_url — which is often a proxy (like a LiteLLM gateway) with no /audio/transcriptions route. So with an OpenAI key present, voice transcribes against api.openai.com out of the box. Fully overridable via stt.<provider>.{base_url,api_key,model} or voice.stt_{base_url,api_key,model}.

2. Configure the transcription model in-app

Settings → Voice now renders the STT provider + per-provider transcription model, backed by:

• a focused /api/config/schema (stt.enabled, stt.provider select, stt.openai.model, stt.groq.model, voice.auto_tts), and
• PUT /api/config, which serve was missing entirely. It deep-merges the renderer's (secret-redacted) draft over the stored config so api_keys and the env block survive, and unwraps saveClawCodexConfig's {config:{…}} envelope. This is what makes all in-app settings edits persist, not just voice.

Model resolution ladder: stt.<provider>.model → stt.model → voice.stt_model → provider default (whisper-1 / whisper-large-v3). The failure message now points at Settings → Voice.

Verified live

The Voice panel renders the STT provider + model fields (with the schema descriptions); saving persists to config and survives a reload with api_keys/env intact; STT resolves to https://api.openai.com/v1 with the openai key. 65 desktop tests green (5 new: canonical endpoint, explicit stt override, model-resolution ladder, PUT deep-merge preserving secrets, schema exposes the model field).

To actually transcribe, put a real OpenAI API key in the openai provider (or set stt.openai.api_key); the chat gateway key won't authenticate against api.openai.com.

🤖 Generated with Claude Code