nx_native_matmul_bench.nx

nx_native_matmul_bench.nx -- matmul-shaped native micro-bench. Most LLM-relevant micro-bench in the paired suite. llama.cpp's transformer forward is dominated by GEMM kernels; this measures how Nishi's native codegen handles the same access pattern at a tractable size. Workload: C = A * B for 128x128 i64 matrices. ~2M multiply-adds. Returns (C[0][0] & 0xFF) as exit code. Compares against gcc -O0 (no opt) and gcc -O2 (auto-vectorized matmul) on the same machine for paired comparison.