A team calls Claude Haiku 4.5 with a system prompt and tool definitions totaling about 2,800 tokens, and places cache_control on the last system block. Every call succeeds, but the usage object always shows cache_creation_input_tokens of 0 and cache_read_input_tokens of 0, and the cost never drops. Requests are sent several minutes apart from a single worker. What is the most likely cause and fix?