nx_q4k_fused_vs_x4_gate.nx

nx_q4k_fused_vs_x4_gate.nx -- THE HEAD-TO-HEAD the SOTA decision actually rests on: the FUSED INTEGER dequant-dot (nx_q4k_dot_row_col) against the LIVE forward kernel (fq4m_rows_x4 / nx_f32_q4k_matmul_x4), in ONE binary, ONE shape, ONE thread. WHY THIS GATE EXISTS. nx_q4k_speed_bench reports the fused integer route beating f32 by ~20-48x, but its f32 leg is SINGLE-THREADED SCALAR EMULATED f32 -- which is NOT the live path. The live path is nx_f32_q4k_matmul_pool_x4: hardware __f32x4_dot AND multicore. Quoting that 20-48x as the shipped win would repeat the exact "10.25x vs SERIAL scalar" error already caught once in this lane. A speedup is only real against WHAT IS ACTUALLY DISPATCHED. THREADING IS DELIBERATELY EXCLUDED, AND THAT IS SOUND. Both routes band-parallelise over the SAME axis (output columns j) with identical banding, so a pool multiplies both sides equally and the single-thread ratio carries over. It is also CONSERVATIVE for the fused side: fused reads only the packed Q4_K bytes while x4 materialises k f32 values per column, so under real multicore memory pressure fused should scale strictly BETTER, never worse. Any win measured here is a LOWER BOUND. NUMERIC REGIME (why bit-exact comparison is legitimate here, and why that is NOT a general claim). Weights are synthetic Q4_K super-blocks with d=1.0, dmin=0, all scales=1, mins=0, so every dequantised weight is an exact integer 0..15. The activation column holds small exact integers. Then: x4 computes an exact-integer f32 sum (|dot| <= 15*3*1024 = 46080 << 2^24, so f32 add is EXACT). fused computes d1=2^24, m1=0, so v = 2^24*q4 (Q24); v*col_q10 is Q34; >>24 recovers Q10 exactly. Both therefore represent the SAME integer and must agree BIT-EXACTLY. On REAL model weights they will NOT -- fused is activation quantisation (W4A-fixed), a genuine precision change. This gate proves the kernels compute the same function; it does NOT license dispatch on real weights without a separate fidelity tooth (nx_q4k_ggml_kat's 2% band is the existing instrument for that). genealogy_id: nx_q4k_x8_gate (data construction, exact regime, refutation-tooth discipline)