Reproducible local-LLM serving recipes (llama.cpp / vLLM) with measured results on multi-3090 hardware