KV-cache decode cost: the real math behind LLM serving bills

A line-by-line audit of how KV-cache growth drives decode-phase cost per token, with batch effects and context-length penalties. For teams serving LLMs in production.