[zhuqiweigit]

Description

In single-controller distillation, teacher scoring localizes rollout RTensors into each teacher worker's _fetch_buffer, but step cleanup omits the teacher. Actor cleanup deletes the source shards without releasing these process-local copies, so teacher memory grows across training steps.

Include the single teacher in step cleanup for both training and rollout teachers. Add RolloutController.clear_batches() to send deduplicated shard IDs to every worker without refetching inputs or deleting source storage. Wait for all worker drains before propagating a failure. CPU regressions exercise repeated training steps with real RTensor serialization, storage, and worker cache eviction.

Related Issue

Follow-up to #1209, covering the single-teacher consumer cache.

Type of Change

• 🐛 Bug fix

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new regression tests added
• Documentation updated (not applicable: internal cache lifecycle fix)
• Branch is up to date with main
• Self-reviewed, including an infrastructure expert review
• This PR was created by a coding agent using the create-pr skill
• This PR is a breaking change

Additional Context

Validation: 163 passed, 1 deselected (the slow model-serving integration test); all pre-commit checks passed.

pre-commit run --all-files
HF_HUB_OFFLINE=1 NO_PROXY='*' no_proxy='*' PYTHONPATH="$PWD" pytest -q \
  tests/test_teacher_rtensor_cleanup.py \
  tests/test_eval_rtensor_cleanup.py \
  tests/test_trainer_utils.py \
  tests/test_train_controller.py \
  tests/test_trainer_eval_before_train.py \
  tests/test_rollout_controller.py \
  -m 'not slow'

The new repeated-training regression fails on the unmodified base because teacher input caches survive the first step. Test environment flags use the cached model files and keep local callback requests out of the HTTP proxy.

Full model-serving, GPU training, and multi-node integration suites were not run; validation targets the CPU-side RTensor lifecycle and controller cleanup paths without launching a training deployment.