OpenAI STT model is hardcoded to whisper-1 — make it configurable

[jason-shen]

Problem

The OpenAI STT client hardcodes the model:

// internal/stt/openai.go
const (
	whisperModel   = "whisper-1"
	...
)

[openai] model exists in config but is only consumed by the LLM client, so there is no way to select OpenAI's newer transcription models — gpt-4o-transcribe and gpt-4o-mini-transcribe — which are more accurate than whisper-1 and, in the mini case, cheaper. For a voice runtime the transcription model is one of the highest-leverage quality knobs available, and right now it is welded shut.

Proposed change

Add a dedicated STT model setting rather than overloading [openai] model — the same [openai] block feeds both LLM and STT, and a single model key cannot mean gpt-4o-mini for chat and gpt-4o-transcribe for transcription at the same time.

1. internal/config/config.go — add STTModel string \toml:"stt_model"`toOpenAIConfig, mirroring how DeepgramConfigcarries bothmodelandtts_model`.
2. Default it to whisper-1 in Load() via setDefault, so behaviour is unchanged for existing configs.
3. internal/stt/openai.go — replace the whisperModel constant with a field on openaiClient, plumbed through NewOpenAIClient.
4. internal/stt/stt.go — pass cfg.OpenAI.STTModel at the case "openai": call site.
5. config.toml.example and the README provider table — document the accepted values.

Worth checking while you're in there

gpt-4o-transcribe and whisper-1 are both batch endpoints, so the existing buffer-until-silence approach in this file still applies unchanged. No VAD or streaming work is needed for this issue — keep the scope to model selection.

Acceptance criteria

• [openai] stt_model = "gpt-4o-transcribe" is used for transcription requests.
• A config with no stt_model still transcribes with whisper-1.
• Config test covers the default and the override.

Pointers

• internal/stt/openai.go — whisperModel const ~L16, NewOpenAIClient ~L56
• internal/stt/stt.go — NewClient, case "openai":
• internal/config/config.go — OpenAIConfig ~L206, Load() defaults ~L300