nx_genweights.nx

nx_genweights.nx -- read model weights straight from a GGUF. The engine's weight source. Until now this lane read weights from fixtures the oracle tap had dumped. That was right for building a RULER -- the tap's f32 conversion is ggml's own dequantization, which is exactly what a differential test needs. It is wrong for an ENGINE: it means the sovereign path cannot run without first booting a third-party binary to export its weights, and it does not scale (one Z-Image layer is ~190MB of raw blocks; 36 layers is ~7GB of fixtures nobody should be dumping). ★ A RULER MAY DEPEND ON THE INCUMBENT. A RUNTIME MAY NOT. The GGUF's tensor names turn out to be EXACTLY the names sd.cpp uses internally ("model.diffusion_model.layers.0.attention.qkv.weight"), so this is a drop-in swap for the fixture loader -- verified by running the same block both ways and comparing. The file is mmapped whole (sys_map_file), not read: a 6.2GB model is paged in on demand and only the tensors actually touched ever reach RAM. HANDLE LAYOUT (i64 slots): [0]=mapped base, [1]=file length, [2]=*NxGgufHeader license_tier: ORIGINAL