nx_ng_radiance.nx

nx_ng_radiance.nx -- CAP-RADIANCE-FIELD, generation 8: a tiny NeRF. The defining recipe of NeRF/instant-ngp: a COORDINATE NETWORK (MLP queried at sample positions) whose outputs are VOLUME-RENDERED into observations, trained by INVERSE-RENDERING loss to recover the field from rendered measurements. Composes the two rungs just built -- the autograd gradient engine (nx_nofloat_autograd) and the volume compositor (nx_ng_volrender) -- on ONE Q16 integer tape: the composite C = sum_i w_i*f(x_i) is nfa_cmul (scale by the fixed transmittance weight) + nfa_vadd, so the WHOLE chain (coord-MLP -> volume render -> loss) backprops end-to-end. No float. Honest gated proof (HARD evidence): a shared MLP f(x)=W2*relu(W1*x+b1)+b2 is applied at N sample coords -> colors c_i; R observations C_r = sum_i w[r][i]*c_i (volume composite, weights = a fixed invertible transmittance matrix) are matched to targets generated from a TRUE field. Gate: T1 LEARNS -- inverse-rendering SSE loss collapses (the gradient flows MLP<-render<-loss), T2 FIELD RECOVERED -- the trained net reproduces the true field at the samples (|f(x_i)-c*_i| small): teeth = it learned the RIGHT field through the renderer, not just any loss drop, T3 BIT-EXACT -- train twice from the same init -> identical integer weights (structural determinism), T4 UNTRAINED -- the initial loss is genuinely large (not vacuous). Sovereign: imports nx_nofloat_autograd + nx_syscalls (no float). license_tier: ORIGINAL expect_exit: 0