nx_roofline_test.nx

nx_roofline_test.nx -- the team DIAGNOSES the real bottleneck of the targets (llama.cpp / ggml), and proves the intelligent lever. Three checks, exit 0 only if all hold: (1) the Q4 quantized MATVEC of single-token decode is MEMORY-bound on a real consumer CPU -- matching the measured reality (AVX2 micro-opt = +0.8%). So SIMD is the WRONG lever. (2) the SAME math, but a batched/reused matmul (each weight reused many times), is COMPUTE- bound -- so the search-governor/superopt IS the right lever there. The roofline tells the team WHICH regime it is in, per kernel, per chip. (3) the data-movement lever (Q4 -> Q2, half the weight bytes) gives ~1.8x in the memory-bound regime -- ~225x more than the SIMD micro-opt. "mental math" (move less) >> brute force. license_tier: ORIGINAL