RL training

Longer term, we need to run RL experiments on minimization. This will require two things: 1. a good base policy 2. good RL code. We have to wait a bit for both.

The issue names no files, tests, or entry points. First clarify the minimization experiments, the required base policy, and the intended RL scope; then identify where training code should live. Done should be defined by an agreed experiment plan and measurable results.

由索引模型根据 Issue 内容生成。