nx_q8_matmul_micro.nx

nx_q8_matmul_micro.nx -- DECISIVE micro-probe of the REAL production Q8_0 matmul path (nx_f32_lazy_weight_new_q8_0 + nx_f32_lazy_matmul -> _lw_q8_0_matmul). The forward's MATMUL bucket is 72% of a decode token (462ms), yet the ISOLATED nx_q8_0_simd_gate clocks this same kernel at ~20 GFLOP/s (=> lm_head should be ~14ms, but the forward pays ~143ms): a 10x in-situ gap. This probe localizes it: WARM same weight, m=1, matmul in a tight REP loop (weights become cache-hot after pass 1) -> per-call us + GFLOP/s + bytes/s. COLD N DISTINCT weight buffers (aggregate >> LLC), one matmul each, NEVER reused -> the real-forward access pattern (every weight read exactly once from DRAM). If WARM ~ 20 GFLOP/s and COLD << that -> the forward is COLD-WEIGHT- STREAMING bound (fix = fewer bytes or batch tokens, NOT a wider kernel). If WARM is ALSO slow -> per-call overhead (mmap/pack/dispatch) or the kernel itself. Shapes: W_gate (k=896,n=4864) + lm_head band (k=896, n=32768). license_tier: ORIGINAL expect_exit: 0