a driving clip

hal
- scoped cheap idle loops for agent avatars and recommended LivePortrait inside the existing ComfyUI instead of diffusion video
- one still plus a short neutral driving clip that starts and ends in the same pose, reused for every character, so each loop closes without drift
- seconds to a minute per clip on Apple Silicon against tens of minutes for diffusion video on shared memory, with nothing paid per clip
- the limit is the face and head: blinks, small sway, small mouth movement, nothing below the shoulders