nx_f32_matmul_block_gate.nx

nx_f32_matmul_block_gate.nx -- referee for the 4x4 register-blocked matmul_t (nx_f32_matmul_t.mmt_block, 2026-09-15). The blocked kernel claims BIT-EXACT equality with mmt_range (the serial reference every other matmul_t path is judged against) on every shape: the interior 4x4 blocks, the ragged right edge (n not a multiple of 4), the ragged bottom edge (m not a multiple of 4), both edges at once, m=1 decode rows, k below and above the block, and the cross-encoder's own shapes (T x 384 x 1536, T x 1536 x 384, T x 384 x 384). Equality is asserted on every output cell of every shape -- a tolerance would be the wrong oracle for a change that only re-interleaves independent sums. The fills are a seeded integer LCG mapped to small f32 values so every product and sum is reproducible on any host. The gate also TIMES the reference and the blocked kernel on the cross-encoder FFN shape and prints both as values (never a threshold: a storming host would make a speed tooth flaky, and the number is the evidence the plan row cites). license_tier: ORIGINAL