nx_qlayer.nx

nx_qlayer.nx -- the full decode LAYER: y = W . x, the operation an LLM does thousands of times per token. Built honestly for the memory-bound regime: the weight matrix is PRE-QUANTIZED ONCE (offline), and the HOT kernel streams ONLY the packed codes + per-block scales from DRAM -- the exact deployment structure of ggml (weights quantized at load, dequant+dot in the inner loop). This is the layer that, moving 3.5x less data at bounded quality, beats llama.cpp where it is actually bottlenecked. The ENGINEER verifies the WHOLE layer output 1:1 vs the exact layer, and the max per-row error is measured against the Council quality floor. license_tier: ORIGINAL