The code in as_completed() waits on a FIRST_COMPLETED event, which means that after the first future completes, it will no longer wait at all. The proposed patch adds a _AsCompletedWaiter and uses a lock to swap out the finished list and reset the event. This is a difficult problem to create a test case for without adding intrusive code to as_completed() to count how many times it loops or create a mock Event object that counts it, so I have not proposed a test.
Good catch. I suppose it was inevitable when I transferred my changes into a SVN checkout that I missed something. Nevertheless, your changeset in r86491 matches my own. Thanks!
