nx_blockfloat_block_gate.nx

nx_blockfloat_block_gate.nx -- BLOCK-FLOAT weights through the REAL no-float TRANSFORMER BLOCK. Composes the gradcheck-verified pre-norm block forward (blk_fwd/blk_out, from nx_nofloat_block_gate: RMSNorm -> Q/K/V -> RoPE -> scaled causal-softmax attention -> out-proj -> residual -> SwiGLU-FFN -> residual) UNCHANGED, but runs it with the weight matrices (Wq,Wk,Wv,Wo,Wg,Wu,Wd) BLOCK-FLOAT quantized (per-row-per-block power-of-2 scale). Proves a real transformer component runs on block-float, and that block-float weights beat per-tensor INT8 at the BLOCK output level while staying bit-exact deterministic. Mixed-magnitude weights (outlier rows) make the range matter. 1 the real block RUNS with block-float weights (non-trivial output) 2 EXCEED (MEASURED): block output error vs full-precision < per-tensor INT8 weights 3 DETERMINISTIC: the block-float block run twice == bit-identical 4 block-float weights are REAL (block output != full-precision output) expect_exit: 0 license_tier: ORIGINAL