nx_gen_rmsnorm_verify.nx

nx_gen_rmsnorm_verify.nx -- SOVEREIGN RMSNorm, verified vs the oracle. out[t][i] = x[t][i] / sqrt(mean_i(x[t]^2) + eps) * w[i] RMSNorm appears four times per DiT block and in every Llama-class text encoder, so this is one of the highest-multiplicity ops in the whole gen stack. Named for the OP; the model, the tap names and eps are all arguments. Usage: nx_gen_rmsnorm_verify <model_id> <x_name> <w_name> <y_name> [rows] [eps_recip] eps_recip eps expressed as 1/N (default 1000000 = 1e-6, the ggml RMSNorm default). Passed as an integer reciprocal because eps is always a negative power of ten in practice and an argv float parser is a defect surface this organ does not need. ONE THING TO KNOW ABOUT THE ORACLE FIXTURE: in this engine RMSNorm ends with ggml_mul_inplace, so its result is a VIEW of the rms_norm output. Tapping a view without protecting its storage dumps recycled memory -- see reference-inplace-view-tap-defect-2026-08-06. The tap now flags the whole view chain, which is why these fixtures are trustworthy at all. Hot loop uses the HARDWARE __f32_* intrinsics, not the nx_f32_* software IEEE-754 twins: the software pair measured 9-10x slower in this lane's matmul and is the single easiest way to make a correct sovereign kernel look like a failed one. license_tier: ORIGINAL