Takeover Call

Operator takes over an in-progress agent-handled call.

Effects:

1. voice_sessions.metadata.mode flips from 'agent' to 'user_call'. taken_over_at + taken_over_by_user_id recorded for audit. voice_sessions.user_id reassigned to the taking-over operator so /rejoin accepts their publish-token mint. 1b. With tab_id, the clicking tab takes the microphone by force BEFORE anything fans out. Every inbox tab reacts to the mode_changed event below by mounting RejoinUserCallOverlay, and each /rejoin claims passively — so without this the first tab to answer won, and live 2026-08-21 that was a background tab: the operator who clicked got listen-only and a "Call is in another tab" chip on their own take-over. Same rule the dispatch applies for Call (internal_voice.py perform_dispatch_room): the acting tab wins, period.
2. PUBLISH on voice:call:control:{call_id} with {"type":"agent_state","disabled":true} so the running worker's _control_subscriber sets _taken_over=true (silences agent's tts_node + llm_node + transcription_node).
3. Frontend WS event voice:call:mode_changed fans out so all inbox tabs flip the active-call header to the user_call layout and RejoinUserCallOverlay auto-mounts.

Worker is NOT re-dispatched. Agent process keeps running silenced for the rest of the call (cost: one idle worker per taken-over call — acceptable for voice-with-zero-users today). Take Over is one-way: operator owns the call until hangup. No "Resume agent" gesture — drop the call and let a new one in if you need agent.

Idempotent: re-clicking on an already-taken-over call re-asserts state (no noop guard on metadata.agent_disabled because that field is no longer written; mode==='user_call' is the truth).