nx_q8_coldwarm.nx

nx_q8_coldwarm.nx -- decides the post-futex decode-matmul direction: COMPUTE-bound (kernel work: AVX2/multi-acc pays) vs COLD-STREAMING-bound (fewer bytes / batch tokens pays). Measures the SAME production Q8_0 matmul two ways at FULL-MODEL working-set scale (~209MB, like the real forward's 24 distinct layers read once/token): WARM = one weight buffer, REPS reps (cache-resident after pass 1). COLD = NB DISTINCT buffers (aggregate >> LLC), one matmul each, never reused (the forward's true access pattern). COLD/WARM ~1 -> compute-bound (kernel is the lever). COLD/WARM >>1 -> cold DRAM streaming (fewer bytes / batch is the lever). license_tier: ORIGINAL expect_exit: 0