nx_q4k_linear.nx

nx_q4k_linear.nx -- the integer GEMM linear: real Q4_K weights x Q10 activation -> Q10 output. sd-server -> Nishi migration (CPU speed path). The integer analogue of qw_linear (nx_f32_qwen_layer): out[t][o] = sum_i act[t][i] * W[o][i], with W a real Q4_K tensor (ggml-correct dequant, ~26x faster than emulated f32) and the activation in Q10 fixed-point. Each output neuron o is one fused nx_q4k_dot_row_col over weight row o (byte stride = n_blocks*144). This is the reusable matmul brick for an integer Qwen layer. Gate proves it matches the f32 path across MULTIPLE real output neurons on blk.0.attn_q. license_tier: ORIGINAL