I have just updated my BOINC client from 5.8.16 to 5.10.7 because this appears to the new recommended production release. Since this update during reporting times the application is now Aborting many of my unstarted units to the extent it nearly deleted the whole of my 80 unit cache.
...
It's a BOINC Scheduler feature which the project recently enabled. If the WU has been finished, reported, and validated for the first two results returned and not yet started by the third, that third host is directed to abort the work and get something useful. The feature only works with BOINC clients 5.8.17 and later.
The change has nothing to do with whether you're running an optimized app or stock, it's all in BOINC. You have probably already gotten this info from the S@H forums where it has been discussed a lot lately, but I thought it would be best to put an answer here in case someone else comes looking.
Joe