nx_kquant_test.nx

nx_kquant_test.nx -- prove the QUALITY lever: finer-grained scales (the K-quant idea) cut the quantization error on REALISTIC weights -- concentrated near zero with occasional OUTLIERS, the real NN case. A naive block of 32 shares one scale, so one outlier inflates the step and blurs the other 31 weights; sub-blocks of 8 isolate the outlier. Exit 0 only if: (1) Q4 @ blk=8 has materially LOWER L2 error than Q4 @ blk=32 (the quality win is real); (2) the cost is honest: blk=8 moves MORE bytes than blk=32 but still far less than fp16; (3) so this is a real point on the quality/data frontier the Council balances. license_tier: ORIGINAL