Add AR Wiener Storm example

[willeastcott]

A webcam AR example that pelts your MediaPipe-tracked head with floppy, physically simulated raw wieners. Whimsical on the surface, but it contributes several reusable pieces to the webcam AR script family.

What's in it

The example (ar-wiener-storm.html): wieners lob in from the hemisphere between your head and the screen on drag-aware ballistic arcs, bounce off an invisible head proxy, cast real shadows onto your face, vanish behind your head, and can be dodged — or batted away — because the physics world is anchored to the room, not your head. The HUD is a dignity meter (drains per hit, seeps back while you dodge) plus a deadpan announcer, driven entirely by app events. ?sim runs without a webcam, ?debug renders the head proxy for fit checks.

New reusable scripts:

• tracked-head.mjs — face tracking with the frame flipped: the camera stays pinned at the origin of MediaPipe's camera space and the tracked head pose drives an entity instead. World-simulated content becomes dodgeable, where the base class's face-locked frame would drag it along with the head. Includes the nose-bridge anchor pin in the flipped frame. Builds on faceTracking, whose smoothed pose and sim orbit are promoted from private to protected.
• head-shadow-catcher.mjs — grounds virtual objects on the real head: an invisible proxy that renders only the shadows cast onto it. The stock engine shadow catcher darkens the canvas multiplicatively, which cannot affect a DOM video behind a transparent canvas, so this one routes the engine's shadowCatcher accumulator into the canvas alpha as premultiplied black (GLSL + WGSL chunk overrides), gated by N·L so shadows never wrap past the terminator. It doubles as a depth occluder. Accepts any model as the catcher surface, with a cranium ellipsoid fallback.
• video-ibl.mjs — low-frequency image-based lighting estimated from the live camera feed (in the spirit of mobile AR lighting estimation), so virtual objects pick up the room's real color temperature and brightness. Rebuilds scene.envAtlas into reused targets on a slow cadence; the declared sky remains the fallback when no camera runs.

New assets: wiener.glb (procedurally authored in Blender: bent capsule, six-bone chain, slick raw-pork material), canonical-head.glb (MediaPipe's canonical face mesh extended into a full cranium, so it registers with the tracked face and shadows bend around the nose and brow — Apache 2.0, sidecar included), and wiener-slap.mp3 (a wet slap, velocity-scaled at playback).

Reviewer notes

• Each wiener is a chain of six capsule rigid bodies linked by 6dof spring joints, skinned one-to-one to the mesh's bones — the bending is physics, not animation. Physics steps at 180 Hz with raised solver iterations (both restored on destroy): head impacts spin the light segments past their joint limits within a single 60 Hz step otherwise.
• Directional lights emit along the entity's negative Y axis, not forward — the light angles in the markup were read back from lookAt + rotateLocal(90, 0, 0).
• The head proxy stays a pc-entity (model nested as a child) because entity: references resolve bare names against pc-entity elements only; a library-side follow-up for that is tracked separately.
• face-tracking.mjs changes are doc-visibility only (three members promoted to protected); no behavior change, and all existing examples using it are unaffected.

Verified in ?sim on both WebGL2 and WebGPU (the catcher's chunk overrides ship in both shader languages); ESLint and the full test suite (1023 tests) pass.

🤖 Generated with Claude Code