LLM API Cost Structure for Agent Fleets: Per-Cycle Economics, Cache/Routing ROI, and Cost-Aware Reasoning Patterns