Control Signal Renderer

Getting consistent structural conditioning -- camera, room geometry, figure count, poses -- into an AI video or image pipeline usually means standing up a real 3D pipeline: Blender or Unreal, a headless-GL context that is notoriously flaky on Apple Silicon, a mesh library, an artist to build sets. This package skips all of it. Every figure is a handful of capsules and spheres, every surface is a closed-form ray intersection (plane, sphere, capsule -- no mesh anywhere), and four pixel-aligned control passes fall out of one shared camera and one shared per-figure rig.

WHAT SHIPS: beauty, depth, segmentation and pose-skeleton frame sequences from one SceneDescription; deterministic placement (a pure hash of the figure's own id -- same input, same output, forever); ten built-in animation states (walk, sit, gesture-talk, point, mop and more) layered on a breathing / weight-shift idle so a multi-second clip never reads as a still with limbs stapled on; an optional per-figure PhysicalityHint (gait pace, gesture amplitude, posture lean, stance width); a 15-joint procedural rig; a typed SceneDescription contract any upstream source can target; a CLI (npm run render ships a runnable synthetic example) and a library entry (buildSceneDescription + renderSequence). Output is plain PNG sequences, optionally assembled to mp4 when ffmpeg is on your PATH. One native dependency (sharp, prebuilt binaries).

The bundled example scene is synthetic. Nothing in this package references any production show.

VERIFIED: 57 tests, measured by running the suite.

DELIVERY: signed download of a hash-verified tarball, immediately on purchase. Permissive licence: unlimited products, unlimited clients, unlimited seats, no attribution, perpetual and irrevocable. One restriction, do not republish the source as source.