On master 872eb6a3a5497721f53ebe82c0da9a3ba3414c13, a failure in pin_modules() can leave successful VBAR faults pinned after cleanup_prefetch_queues(). The queue still contains the raw upcoming module, which cleanup skips. A nonresident module that fails preparation may also lack its _prefetch cleanup record.

A model-free CPU regression module is available here. To reproduce in an isolated upstream checkout at the above commit, copy that test file into tests-unit/comfy_test/ and run:

python -m pytest -q tests-unit/comfy_test/test_prefetch_recovery.py

Result: 4 failed, 4 passed. In test_failed_prefetch_keeps_all_faulted_modules_reachable, two VBARs fault successfully and preparation of the second module raises. Terminal cleanup makes no vbar_unpin calls; expected behavior is one unpin per successfully faulted VBAR and no remaining _prefetch records. Other cases cover failure during the batch and failure in real pin_modules() after casting.

Registering the batch before pinning and publishing minimal batched cleanup state before preparation makes all 8 tests pass. The original exception still propagates.

These tests inject errors and mock device boundaries; no model files or GPU workload are needed. This reproduces Python cleanup behavior, not a particular driver crash or live-model memory-pressure threshold.