Make the `model` field functional in Realtime API

[chen1plus]

Currently, the model field in the Realtime API acts as a static placeholder.

[ivan-digital]

Thanks for filing this @chen1plus — confirmed, model is hardcoded on session.created and ignored on session.update (Sources/AudioServer/AudioServer.swift:272, 290-301).

Plan: accept model in session.update and use it to pick the local engine by the same names the CLI already uses — qwen3, parakeet, nemotron, omnilingual for ASR; qwen3, cosyvoice, voxcpm2, magpie for TTS. Unknown names accepted-and-echoed without changing the backend. Stays OpenAI-shaped, no name aliasing, lets the existing non-OpenAI engine field be deprecated in a follow-up.

Does that match what you had in mind?

[chen1plus]

Yes, but I'd like more detailed options such as qwen3-0.6b, qwen3-0.6b-coreml