Although the test has never been observed to fail on Windows, I think the failure is genuinely a race in the test.
File handles (like events) are signaled in order to mark that an I/O operation has completed. In this case the I/O operation includes manipulating data on both ends of the pipe, and as part of that may signal the other end. Internally, however, the file handle must logically happen *after* all of this processing has taken place, not least because (given the Windows I/O architecture) it is the job of the I/O manager, not the IRP handler.
Since the purpose of the test is probably just to check that the file handle will be signaled after a synchronous I/O operation has completed, we may as well check it from the client thread, after we know for sure that it has.