llama.cpp fork: experimental --override-moe-top-k N (force N routed MoE experts per layer) — proof of concept with benchmarks, perplexity, and graph/kernel evidence