This happened when we got two create cells in a row from the same TLS
connection. It would hand one to the cpuworker, and then immediately
handle the second one -- after it had registered that the first one was
busy, but before it had updated the timestamp that we use to decide how
*long* it's been busy.


svn:r6606