nx_ng_gsplat.nx

nx_ng_gsplat.nx -- CAP-GAUSSIAN-SPLAT, generation 10: 3D Gaussian Splatting (Kerbl et al. 2023, arxiv 2308.04079, Inria/EU). The other pillar of the radiance-field frontier: instead of an MLP queried along rays (NeRF), the scene is EXPLICIT Gaussian primitives that are SPLATTED (each projects to a footprint) and ALPHA-OVER composited front-to-back -- real-time, and differentiable so the primitives can be optimized. This rung REUSES the alpha-over compositor proven in CAP-VOLUME-RENDER (T_g = prod(1-a_h); C = sum T_g*a_g*c_g) and adds the 3DGS-specific pieces: the explicit Gaussian PRIMITIVE + its SPLAT FOOTPRINT (a Gaussian falloff exp(-(x-mu)^2/2sigma^2) via the autograd module's fixed-point nfa_fxexp). 1D image, Q16, no float. Honest gated proof (HARD evidence): N Gaussians (centers mu_g, width sigma) splat onto a P-pixel image; their per-pixel opacities a_g(p)=footprint are alpha-composited (the reused compositor) -> a rendered image. Gate: T1 FORWARD SPLAT -- each Gaussian's footprint PEAKS at its center and falls off (a real splat), T2 DIFFERENTIABLE FIT -- GD on the Gaussian COLORS reconstructs a target render (loss->~0, colors recovered): the explicit-primitive scene is differentiably optimizable (the 3DGS training loop), T3 NEG-CONTROL -- ascending that gradient diverges (the gradient has teeth), T4 BIT-EXACT -- train twice -> identical integer colors. HONEST: 1D, fixed mu/sigma/opacity (only colors optimized); full 3DGS = 3D gaussians with optimized position/covariance/opacity + tile sort + adaptive densification = the follow-on. Sovereign: nx_nofloat_autograd (tape + nfa_fxexp) + syscalls, no float. license_tier: ORIGINAL expect_exit: 0