Sessions: expose turn-detection / server-VAD / barge-in config in @app.session

[AbirAbbas]

Follow-up from #654.

The session config the control plane sends to OpenAI (createOpenAIRealtimeCall in control-plane/internal/handlers/sessions.go) is currently minimal: type, model, instructions, audio.output.voice, tool_choice. There's no way for a session author to configure turn detection / server-side VAD / interruption (barge-in), which are table-stakes for a usable voice UX.

Why it matters

• Without server VAD + turn detection config, the app can't tune when the model considers a turn complete, silence thresholds, or whether the user can interrupt the model mid-utterance. These are the difference between a demo and something people will actually talk to.

Acceptance criteria (behavior)

• @app.session(...) (and the TS/Go equivalents) accept turn-detection / VAD options (e.g. turn_detection type, threshold, silence duration, create_response/interrupt_response) and the control plane forwards them into the OpenAI realtime session config.
• Defaults produce a sane interruptible voice experience (barge-in works out of the box).
• Validation rejects unsupported combinations explicitly (consistent with the existing provider/transport validation philosophy — no silent inference).

Ref: OpenAI realtime session turn_detection config.

[AbirAbbas]

Triage

Classification: feature ✓
Suggested labels: enhancement

Likely affected files:

• ./reasoners/repo_config.py (matched: config, config, config, config, config)
• ./reasoners/models.py (matched: model, model, model)
• ./reasoners/webhook_router.py (matched: out)

---

Triage by github-buddy