I tested LLM routing to understand why switching models mid-session can erase the savings promised by cheaper inference