Reserve KV Cache for the Output You Promised, Not Just the Prompt You Received

Admission control that only measures the prompt lets short requests with big generation caps OOM your LLM server. Reserve for output tokens.