nx_q4k_fused_fidelity_gate.nx

nx_q4k_fused_fidelity_gate.nx -- DOES THE FUSED INTEGER GEMM TELL THE TRUTH ON REAL TRAINED WEIGHTS? THE QUESTION THIS ANSWERS, AND WHY SPEED CANNOT ANSWER IT. nx_q4k_fused_vs_x4_gate proves the fused kernel is >=9.9x faster and BIT-EXACT -- but only in a synthetic regime built from d=1.0, dmin=0, scales=1 and small integer activations, where every value is exactly representable. That regime is chosen to isolate the KERNEL's logic. It says nothing about REAL weights, because the fused route is not a reassociation: it is ACTIVATION QUANTISATION (f32 -> Q10). Dispatching on the strength of a speed gate would ship fast wrong answers -- the speed bench in this very lane stayed GREEN on a deliberately corrupted kernel. SO THIS GATE IS ADVERSARIAL BY CONSTRUCTION: * REAL weights -- the actual blk.0.attn_q.weight Q4_K super-blocks from the live Qwen3-4B model, with their real trained super-scales and 6-bit sub-scales/mins (nothing synthetic). * Activations DELIBERATELY NOT Q10-REPRESENTABLE. Values are built and then multiplied by the f32 constant 1/3 (0x3EAAAAAB), which has an infinite binary expansion, so every activation carries bits BELOW Q10's 1/1024 resolution. Feeding Q10-exact activations would have flattered the fused path and measured nothing -- the quantisation error would have been zero by construction. * The reference is nx_f32_q4k_matmul_x4: the kernel actually dispatched today, not a strawman. VERDICT UNIT: error is reported in PER-1024 of the reference magnitude (1024 = 100%, 10 ~= 1%), so the actual number is visible and reviewable rather than hidden behind a pass/fail. genealogy_id: nx_q4k_speed_bench (real-model load) + nx_q4k_ggml_kat (2% tolerance discipline)