P4B: Live call quality — latency and memory leaks

[joelteply]

Memory not freed after live chat, STT/LLM/TTS latency, simultaneous speech. See P4B.

[joelteply]

Avatar camera geometry bug in live view:

Most avatars show the BACK of their heads — camera is behind them. Some face correctly (Grok, Together, Gemini, Helper AI, Claude), others are 180° rotated (Groq Lightning, Local Assistant, Teacher AI, Fireworks AI, GPT, DeepSeek).

Likely cause: inconsistent VRM model facing direction. Some models export facing +Z, others -Z. The per-slot camera setup doesn't compensate.

Fix options:

1. Detect model facing direction at load time and flip camera for reversed models
2. Rotate the avatar model to face the camera consistently
3. Store facing direction in the avatar catalog and apply per-model correction

Screenshot captured in docs/images/bug.png

[joelteply]

CORRECTION: Not camera geometry — vertex/mesh destruction.

The avatars aren't backwards. The vertices are EXPLODED. Hair stretched into spikes, faces shredded, geometry destroyed. Compare:

• Working: Grok, Together Assistant, Gemini, Helper AI, Claude Assistant — clean geometry
• Broken: Groq Lightning, Local Assistant, Teacher AI, Fireworks AI, GPT, DeepSeek — mesh torn apart

Likely causes:

1. VRM joint limit: Bevy 0.18 can't handle >128 joints. Models with more joints get garbled skinning.
2. Bone index overflow: Vertex bone indices reference bones that don't exist → vertices fly to origin or infinity
3. VRM version mismatch: Working models may be VRM 0.x, broken ones VRM 1.0 with different bone structure
4. 100Avatars vs VRoid: The working ones are the 8 VRoid CC0 models we kept. The broken ones might be 100Avatars models that weren't cleaned up, or the avatar assignment is wrong.

This is a rendering bug, not a camera bug. The Bevy skinned mesh pipeline is corrupting vertex positions for certain models.

[joelteply]

Superseded: Memory leak tracked in #603 (4GB guard added). STT latency fixed (#599-601). Simultaneous speech handled by LiveKit active speaker detection.