nx_ng_avatar.nx

nx_ng_avatar.nx -- CAP-NEURAL-AVATAR, generation 13: animatable neural humans / gaussian avatars (the Samsung-Moscow neural-avatar lineage, src ng_region_ru_neuralavatars + gaussian-avatar work). The defining property past a STATIC 3D scene (3DGS/NeRF): one CANONICAL set of primitives is DEFORMED by a pose/expression parameter (a deformation field maps canonical -> posed) so a SINGLE learned model renders MANY poses -- and a NOVEL pose (not trained) is produced by INTERPOLATING the same deformation = generalization, not memorization. Reuses the splat machinery: a canonical gaussian has center mu0; pose theta moves it mu(theta)=mu0 + theta*db (a learnable per-gaussian deform basis db). Render = splat at the posed center (CAP-GS-REALTIME footprint) into a 1D image; fit the deform basis from a few posed target frames. Q16, no float. Honest gated proof: T1 ANIMATABLE -- after fitting db from training poses, the model REPRODUCES those poses (loss->~0). T2 NOVEL-POSE GENERALIZATION -- a pose theta NOT in training renders correctly (the gaussian lands at the right place) because the deform is linear in theta = interpolation, not a stored frame (the load-bearing avatar property). T3 NEG-CONTROL -- a STATIC model (no deform, db=0) CANNOT match the posed frames (high loss): the deformation is what's doing the work. T4 BIT-EXACT. HONEST: 1D, 1-DOF linear deform, fixed footprint (full avatar = 3D + skeleton/blendshape rig + learned non-linear deform = follow-on). Sovereign: nx_nofloat_autograd (tape + nfa_fxexp footprint) + syscalls. license_tier: ORIGINAL expect_exit: 0