How Over‑Optimized Prompt Caching Kills Real‑World LLM Application Performance