Testing rollout generation currently requires full RolloutWorker infrastructure including JAX/mesh setup, background InferenceServer threads, weight transfer via Arrow Flight, Ray curriculum actors, and rollout storage backends. This creates slow tests with high overhead (100+ lines for simple cases) and complex debugging buried in async workers. We need a lightweight RolloutManager class that provides synchronous rollout generation for unit tests and debugging while maintaining compatibility with production pipeline output formats.

Parent Issue: #1738

Relevant Code

• src/marin/rl/rollout_worker.py - Production worker to extract core logic from
• src/marin/rl/rollout_worker.py#L203-L269 - _sample_batch() method with core rollout generation
• src/marin/rl/environments/base.py#L30-L60 - Environment protocol and LevanterInferenceContext
• tests/rl/integration/config.py#L443-L683 - Current test workarounds to replace
• tests/rl/integration/tasks.py#L41-L138 - Tests manually creating batches