Build step 3: live Kinect capture + ellipsoid proxy from side cameras

The occlusion proxy (PLAN.md core design): splatgen fits a
head/shoulders ellipsoid to each frame's depth cloud (the nearest
cluster; the backdrop is further out), then casts every side-camera
pixel ray onto it and keeps the hits the depth map has no data for —
behind the observed subject surface, or over background. Those become
splat with the ellipsoid's outward normal, the side camera's pixel
colour, inflated thickness, and lower opacity (the plan's lower
confidence weight). The synthetic head carries side tints only its
flanks show, so the recolouring is provable: proxy splats carry the
side cameras' colours, and a GPU test shows the tinted flanks appear
in the rendered view only with the proxy on.

Live capture: the capture crate opens the Kinect v2 (kinect-one: CPU
depth processing, pure-Rust JPEG, factory registration) plus the two
webcams (nokhwa, matched by id or name), runs each on its own thread,
and composes frames timestamp-aligned with ~1-frame skew tolerance.
The source's rig carries the Kinect's factory depth intrinsics.

- rig: shared Ellipsoid geometry; DepthFrame carries side RGB;
  DepthSource exposes the source's full rig
- replay: synthetic scene's head flanks tinted (proxy ground truth);
  source emits side RGB; recording format carries the rig + side files
- splatgen: SubjectFit (ellipsoid + cluster depth range), the
  occlusion rule, and the proxy fill; SplatSource owns the rig
- capture: KinectRigSource (code-complete; hardware verification
  pending — no USB in this container; opening fails fast and clearly
  without a Kinect, which is tested)
- sink: --source synthetic|recording|kinect, --recording-path, and
  proxy knobs join the config surface; main dispatches all sources
- tests: 75 across the workspace, including the recording path
  end-to-end and the proxy's side-colour assertions