SSTable leak after stress and repair

I have a dtest that fails intermittently because of SSTable leaks. The test logic leading to the error is:

• create a 5-node cluster
• insert 5000 records with stress, RF=3 at CL=ONE
• run flush on all nodes
• run repair on a single node.

The leak is detected on a different node than repair was run on.

The failing test is here. The relevant error his here, along with the errors from the entire 5-node cluster. In these logs, the repair was run on node1 and the leak was found on node2.

I can bisect, but I thought I'd get the ball rolling in case someone knows where to look.