nx_q4k_matmul.nx

nx_q4k_matmul.nx -- fused-dequant Q4_K row × column dot product. Phase C foundation brick (conductor cardinal 2026-05-19). Composes nx_dequant_iter.nx with the existing matmul pattern to produce a row × column dot product WITHOUT materializing a full dequantized weight tensor. What the brick does: * Caller holds the Q4_K-quantized weight buffer (e.g. raw bytes from a GGUF file -- composes with nx_gguf_load_lazy) * Caller holds the dense column vector (Q10 i64) * This brick walks the row's Q4_K blocks, dequants each block's 256 values inline, multiplies by the corresponding column entries, and accumulates the sum * NO scratch buffer for dequantized weights; the substrate-side RAM cost is the column vector + per-block iterator (~152 bytes) Honest framing (per no-strawman-perf-comparisons cardinal): - This brick proves the FUSED dequant+dot composition works and produces bit-exact results vs the materialize-then-matmul path. - Per-element throughput on qemu-RV64 is measured by the bench below. This is NOT a claim about CUDA or llama.cpp -- those systems have their own per-element costs that require running them on the same hardware to compare honestly. Output unit: dot product is returned in Q34 (sum of Q24 weight × Q10 column products). The Q4_K weight dequant is now Q24 (EXACT super-scales -- Q10/Q14 underflow/flip sign on real trained-model super-scales). Caller uses nx_q4km_q20_to_q10 to recover Q10 (>>24 round-half). Q34 is kept internally so successive matmuls can compose without intermediate normalization rounding. genealogy_id: ggml_q4k_dot_canon + standard_fused_dequant_pattern lineage_id: substrate_q4k_matmul_v1_fused