ElevenLabs TTS: chunk_length_schedule not forwarded in multi-context websocket init

[IanSteno]

Bug summary

chunk_length_schedule is accepted by the ElevenLabs TTS plugin constructor/options, but in the multi-context websocket path it is not forwarded in the init packet (generation_config.chunk_length_schedule).

This makes the setting effectively a no-op for current releases using the multi-context websocket flow.

Affected area

• livekit-plugins-elevenlabs
• livekit/plugins/elevenlabs/tts.py

Repro/observations

• Pre-regression behavior (tag livekit-agents@1.2.6) sends:
  • init_pkt["generation_config"]["chunk_length_schedule"]
• Current behavior (tag livekit-agents@1.4.3) accepts chunk_length_schedule, but websocket init only sends:
  • text, voice_settings, context_id (plus optional pronunciation dictionaries)
  • no generation_config

Why this is a bug

The plugin API and docs expose chunk_length_schedule, but it is not propagated to the backend request path where it should apply.

Proposed fix

Tracked in PR:

• fix(elevenlabs): restore chunk_length_schedule in WS init payload #5006

That PR:

• restores generation_config.chunk_length_schedule in multi-context websocket init packets
• includes tests for packet composition
• aligns default auto_mode behavior when an explicit chunk_length_schedule is provided

Additional context

This appears to have regressed during the migration to multi-context websocket handling.