nx_nofloat_blockfloat_gate.nx

nx_nofloat_blockfloat_gate.nx -- BLOCK-FLOAT quant rung: per-block power-of-2-scaled integer = MXFP8's dynamic-range fix, but DETERMINISTIC. "Use it for building": grounded in the operator's banked arithmetic (knowledge/research/2026-06-24-arithmetic-grounding-blockfloat.md + knowledge/library/arith_fp8_training.txt -- FP8's case against INT8 is dynamic range; MXFP8's fix is per-block(32) power-of-2 scale in E8M0). Power-of-2 = integer shift = exact + associative, so we get FP8's range advantage WHILE keeping the bit-exact determinism moat. Proves the FORMAT UPGRADE per-tensor INT8 -> per-block float: criteria: 1 block-float round-trips the small block EXACTLY (its block scale e=0): reconstructed == original 2 EXCEED (MEASURED): total reconstruction error block-float < per-tensor on a mixed-magnitude tensor 3 DETERMINISTIC: quantize twice -> bit-identical mantissas + scales (vs float MXFP8 non-determinism) 4 per-tensor scaling DESTROYED the small block (-> all 0); block-float PRESERVED it (the super-weight fix) expect_exit: 0 license_tier: ORIGINAL