nx_matmul_t_pool_gate.nx

nx_matmul_t_pool_gate.nx -- gate + MEASUREMENT for the pooled transposed matmul (nx_f32_matmul_t_pool) that threads lm_head. lm_head is the forward's biggest single matmul: at decode it is nx_f32_matmul_t(A[1,896], B[vocab,896], C[1,vocab], 1, 896, 151936) = 136M MACs, and it ran the SCALAR SERIAL mmt_range every token (fork was disabled in the organ). This gate runs that EXACT shape serial vs pooled so the printed serial_us IS lm_head's real per-token cost, and proves the threaded result bit-identical. BIT-EXACT ON ANY DATA: flat-range banding computes each C[idx] wholly inside one band with the identical accumulation order as serial -- so pool == serial is bit-exact regardless of magnitude. Values are still kept small-int-exact (finite, no NaN) so the bit-compare is clean. Checks (6): 1 small prime shape (m=3, n=37): pool == serial bit-exact (band-boundary correctness -- 37 is not a multiple of 16) 2 serial oracle OK at the real lm_head decode shape 3 pooled result bit-exact vs serial at the real shape 4 auto-worker pool (nx_pool_new(0)) == serial 5 pool reuse (second call, new A) == fresh serial 6 SPEEDUP: pooled >= floor x serial at the real shape (serial_us printed = lm_head's real per-token cost) lineage_id: matmul_t_pool_gate_v1