nx_q4k_real_gemm.nx

nx_q4k_real_gemm.nx -- the CPU SPEED PATH proven on a REAL Qwen weight: integer fused-dequant-dot vs f32. sd-server -> Nishi migration (CPU perf track, SOTA-grounded: llama.cpp/ggml = quantized-integer GEMM, not emulated f32). Reads the real `Huihui-Qwen3-4B-...-Q4_K_M.gguf`, takes blk.0.attn_q.weight (Q4_K), and computes one output-neuron's projection two ways over row 0: (A) INTEGER path -- `nx_q4k_dot_row_col` (fused dequant+dot, integer Q34 = Q24 weight x Q10 col; the FAST path) -> Q10 -> f32. Q24 super-scales are EXACT (Q10/Q14 underflow/flip-sign on real weights). (B) F32 path -- `nx_q4k_to_f32` (dequant to f32) + our emulated-f32 dot (the CORRECTNESS path). Gates: (1) fused Q34 == manual sum of `nx_gguf_dequant_q4_k` (Q24) x activation (Q10), BIT-EXACT (the fused kernel is correct on real weights); (2) integer(Q24)-path f32 ~ f32-path f32 within ~2%, both ~ +0.21576 (the true ggml row-0 dot). Bounded prefix read; only the one row is dequantized. license_tier: ORIGINAL