Accept Pending Invite

Response for 1:1 inbound pickup (telegram_1to1 + twilio_1to1).

Single-shape (no error arm): failure modes raise HTTPException (502 / 409 / 410 etc.) so the body is never returned on error. The frontend's acceptVoiceInvite calls apiClient.post, which throws APIError on any non-2xx — try/catch handles failures uniformly with the outbound /api/voice/user-call flow.

Per the unified-call-path plan (D3), url/token/room are REQUIRED so the frontend can mount <CallContactOverlay> directly without round-tripping through /api/voice/rejoin/{call_id}. This matches StartUserCallResponse's shape, so both flows converge on the same overlay component with identical props.

Response for telegram_group invite pickup — operator brings the bot into the call but does NOT join as a publisher (the group listener Bot is the participant). No browser-side LK creds because the operator listens via the separate <ListenOverlay> flow, which mints its own subscriber token via /api/voice/listen-in/{call_id}.

Kept distinct from AcceptInviteResponse (1:1 unified shape) so each path has a single, statically-checked contract — no Optional shims, no shared discriminator field.

voice_call_thread_channel_ref is the channel_ref of the new kind='voice_call' sub-thread the bot's join creates — frontend uses this verbatim to navigate the inbox to the active-call view. Server-computed because Telegram's int64 tg_call_id loses precision when round-tripped through JS Number, so the frontend can't reliably reconstruct it from the lossy invite payload.