An AI agent wrote a missing Arm kernel and made llama.cpp prompt processing 6.5x faster

Llama.cpp: 6.5x faster prompt processing on Arm with a NEON kernel

llama.cpp's tiled K-quant matmul had inner kernels only for AVX-512 VNNI, AVX2, and AVX, so on Arm, formats like IQ3_M fell back to re-unpacking weights for every dot product. A 164-line NEON SDOT kernel, written by an AI agent, fixes that: IQ3_M jumps from 11.5 to 74.5 tokens/s on Neoverse N1, and other low-bit formats gain 1.6x to 5.9x. Generation speed is unchanged, and the patch hasn't been submitted upstream.