Most LLM inference runtimes have no idea a physical deadline exists. This one refuses admission rather than miss a 33ms robot control cycle, evicts KV cache by meaning instead of age, and is written entirely in hand-written CUDA — no cuBLAS, no libtorch.