nx_f32_groupnorm.nx

nx_f32_groupnorm.nx -- software-f32 Group Normalisation (Wu & He 2018), the R3 vision-op re-tier. sd-server -> Nishi migration, rung 3 ("unify the two numeric universes"): the existing diffusion GroupNorm `nx_groupnorm.nx` is i64 Q10 (10 fractional bits) and has only ever run on identity weights; this is the SAME math in the sovereign software-f32 tier (`nx_f32_*`, bit-exact IEEE-754 emulated in i64) so the REAL dequantized GGUF weights (`nx_q4k_to_f32` etc.) flow through it. GroupNorm(32) is the load-bearing norm in SD/Flux/Z-Image UNet + VAE ResBlocks. Tensor representation: a FLAT `*i64` whose every element is an f32 bit-pattern, laid out [N,C,H,W] row-major, with dims passed explicitly. This deliberately avoids `nx_tensor`'s F32 dtype, which is DTYPE_GATED / unusable today (TODO_COMPILER_F32_F64 parked) -- the flat-f32-bits convention is how the whole `nx_f32_*` stack already works. Composes ONLY gated primitives: nx_f32_add/sub/mul/sqrt (nx_f32.nx), nx_f32_div (nx_f32_div.nx), nx_i32_to_f32 (nx_f32_cvt.nx). No libm, no third-party. eps = 1e-5 (Z-Image norm_eps), built from cvt+div so there is no hand-coded float bit-pattern. Math (per group g of a sample n): mean,var over (C/G * H * W) elements; y = (x-mean)/sqrt(var+eps)*gamma+beta. license_tier: ORIGINAL