Add PPO evaluation and greedy baseline reporting

[chenmaike]

Summary

• refactor environment reward computation and add an evaluate() helper for PPO/greedy runs with optional temperature sampling
• automatically run evaluations after training, log average WSR/DVP/constraint rate, and export comparison to JSON/CSV using the same seeded topology
• overlay a greedy WSR baseline line on the observability plot for visual comparison

Testing

• python -m compileall test.py

---

Codex Task