feat: show face, neck, torso and arms tracking on camera

[Endika]

The overlay drew hands only, so "it is not working" and "it cannot see your face" looked identical. Now every tracked part is drawn in its own colour, with a chip per part that goes green when it is being tracked and red when it is not.

Part	Colour	Source
Manos	violet	HandLandmarker
Cara	cyan	FaceLandmarker
Cuello	pink	Pose, nose → shoulders
Torso	amber	Pose, shoulder line and hip box
Brazos	green	Pose, shoulder → elbow → wrist

Separate colours because the parts fail independently: hands track fine at arm's length while the face is out of frame, or the torso is cropped while everything else is good. One uniform skeleton would hide which.

The face is drawn as points rather than mesh edges — 468 edges are unreadable at video size, and a scatter already shows it is tracked. Legs are not tracked at all: nothing below the hips carries meaning in LSE.

Why pose is here at all

Measured, not assumed. On SWL-LSE's held-out test split, normalising hand position against the torso instead of the image frame is worth +5.7 points of top-1 (0.632 → 0.689). "Hand at chin height" is a fixed number in body coordinates and a moving one in image coordinates — it changes when the signer steps closer.

Pose and face load with .catch(() => null) and the app keeps working without them: hands are the only model recognition genuinely needs, so the other two are enhancements, not dependencies. source.available reports which came up.

Engine grows from ~32 MB to ~41 MB; the storage panel says so and can still free it.