ik_llama and BeeLlama both finish in 22-23s on the am17an 9-prompt harness vs mainline llama.cpp's 37s — 1.66x and 1.62x speedups via opposite strategies.