RolloutWorker currently has hardcoded evaluation logic in its main run loop that executes micro-evaluations every 10 steps and full curriculum evaluations every 100 steps. This tight coupling makes it difficult to customize evaluation behavior, add monitoring hooks, or test the worker without monkey-patching. We need to replace this with a flexible hook/callback system that allows arbitrary hooks to be registered while preserving existing evaluation behavior through a default EvaluationHook.

Parent Issue: #1738

Relevant Code

• Main run loop with hardcoded micro eval
• Main run loop with hardcoded full eval
• _evaluate_lesson method
• _evaluate_curriculum method
• Eval frequency config
• Test monkey-patching workaround