Training a 4B model to produce 81% faster query plans than Postgres

...or how to make Qwen learn query optimization via agentic reinforcement learning