Run one voice function tool for a responses-mode session.

The backend Responses model's function call arrives at the app; the app relays {call_id, session_id, thread_id, arguments} here and returns our response as the function_call_output. ask_every is executed by the app (a full Every agent turn), so it is a 400 here. Disabled delegation or unavailable tool names are 404, including calls from already-minted sessions. Every read and write runs as the caller (see services/voice_tools.py).