nx_roofline.nx

nx_roofline.nx -- the team's "where is the real bottleneck?" intelligence (the mathematician's FIRST move: diagnose the binding constraint before spending effort). Polishing a memory-bound kernel's arithmetic is wasted work -- llama.cpp's AVX2 prefetch on the Q4 dot product bought +0.8% because token generation is MEMORY-BANDWIDTH bound, not compute bound. This module makes that judgment explicit and hardware-specific (it "grows to meet the hardware": ridge point is computed from the chip's own peak flops + bandwidth, the vertical co-design north star). ROOFLINE (Williams/Patterson 2009): arithmetic intensity AI = flops / bytes_moved. The machine has a ridge point R = peak_flops / peak_bandwidth. AI < R => MEMORY-bound (the lever is bytes moved: quantization, sparsity, fusion, reuse -- "mental math", not SIMD). AI > R => COMPUTE- bound (the lever is the inner-loop instruction sequence: superopt / the search governor). Attainable perf = min(peak_flops, AI * peak_bandwidth). All comparisons are EXACT integer cross-multiplications (no float, no division truncation), and peak_flops/peak_bw are taken in GFLOP/s and GB/s so the products stay well inside i64. license_tier: ORIGINAL Refs: Williams, Waterman, Patterson, "Roofline" CACM 2009.