nx_nxa_anim_lib.nx

nx_nxa_anim_lib.nx -- THE ONE NXA CLIP EVALUATOR: channel-2 packed world-delta keys, sampled at a clip time, blended toward an idle POSE, and written as the DUAL-QUATERNION ROWS the cast shader skins with. Integer-only and syscall-free so the wasm engine imports it; every consumer (engine, gates, native players) reads ONE implementation of the format's own runtime form. WHY IT EXISTS (GE23, measured 2026-09-06 on the served beach page): the WebGPU cast pipeline already skins on the GPU from a per-frame joint table, and the ONLY per-frame page logic left was animEval -- hand-written JavaScript sampling ANIM tracks with shortest-path nlerp, subtracting a fitted linear root drift, blending toward the frozen stand pose, then forming T = b + dt - D b and the dual half T x real / 2. That is the format's normative runtime form (knowledge/nxa_format_spec.md: M x = D (x - b) + b + dt, shortest-path lerp between keys) expressed in the page instead of the engine. This lib is that arithmetic in NishiLang, driven by the ENGINE's own stride phase rather than a page clock. THE UNIT TRAP, NAMED (spec + nx_nxa_play's audit): ANIM channel-2 dt LANES are MILLIMETRES packed /100 to fit i16, while SKEL binds and the rows are in 0.01 mm model units, so dt lanes are scaled by NA_DT_MM_TO_MODEL here. POSE entry dt words are ALREADY model units (nx_nxa_pose applies the x100 at freeze time; debt 1785380029) and are NOT scaled again. Two sections, two conventions, both cited. KEY LAYOUT (spec, channel 2, two words per key, little-endian lanes): w0 = [t_ms u16][dqx i16][dqy i16][dqz i16] w1 = [dqw i16][dtx i16][dty i16][dtz i16] ROW LAYOUT (what nxGpuCast uploads, per joint): real = q (q12), dual = (T x q)/2 with T in model units, laid out as out[0..nj*4) real rows then out[nj*4..nj*8) dual rows; the page divides both by 4096. license_tier: ORIGINAL No hw writes (Rule 26).