Ollama-shape inference manager using Tom's TurboQuant llama.cpp. FIFO queue + grace + IDLE_HOT hot-hold + model swap on Blackwell.
