nx_f32_hw_matmul_curve.nx

nx_f32_hw_matmul_curve.nx -- diagnose the matmul bottleneck (compute- vs memory-bound). Times the hardware-float GEMM across growing N (matrices grow from L1- to beyond-L2-resident) and reports MFLOP/s per size. The SHAPE of the curve names the next lever, evidence-first: * FLAT across N -> compute-bound on the scalar f32 MAC -> packed SIMD (addps/mulps, 4-wide) is the ~4x lever (needs the sovereign compiler extension: parse+IR+x86 backend). * DROPS at large N -> memory/cache-bound -> transpose-B + cache blocking is the lever (pure .nx). In-organ clock (sys_now_ms), sovereign. license_tier: ORIGINAL