nx_kquant.nx

nx_kquant.nx -- the QUALITY lever for quantization (quality is slightly more valuable than speed): ggml's K-quants beat naive Q4_0 mainly by FINER-GRAINED SCALES -- a big block shares ONE scale, so a single outlier weight inflates the step and wrecks the precision of every other weight in the block. Smaller sub-blocks ISOLATE outliers: only the sub-block holding the outlier pays, the rest stay sharp. This module makes the block granularity a PARAMETER, so the team can trade a little more scale storage for materially better quality -- another point on the quality/data frontier the Council balances. (Real NN weights are concentrated with occasional outliers, the exact case where this wins.) license_tier: ORIGINAL Refs: ggml Q4_K/Q6_K K-quants.