nx_membw.nx

nx_membw.nx -- resolves the open question: is the decode matmul's ~5 GB/s (nx_q8_coldwarm COLD) a HARD memory-bandwidth cap (=> AVX2/prefetch WON'T help, the ~3.5-7x llama.cpp gap is environmental/WSL) or is the SSE kernel LEAVING BANDWIDTH ON THE TABLE (=> a wider kernel WOULD help)? Measures the RAW achievable sequential-read bandwidth of this process: allocate a >>LLC buffer, commit it, then stream-sum it (hardware prefetcher engaged). If raw ~5 GB/s -> bandwidth-capped (matmul is at the floor). If raw >>5 GB/s -> the matmul kernel is the limiter (AVX2 pays). expect_exit: 0