Chances are you should trust the 2010.2 results. I did a search of the Corrected Defects list, and there are several candidates. Some are caused by changes in defaults which had knock-on effects -- this sort of thing is generally corrected in the Update Releases. I noticed that both 'clean' runs were done with update release versions, while the 2009.3 version was a quarterly release (not an update release).
To pinpoint the specific bug would require a complete testcase with both the entire rule file and the design it was run on. You could try looking through the Corrected Defects document for each intervening release, but as they are only one or two sentences each, it may be hard to identify the exact match.
Thanks for your detail explain!
I will use 2010.2 as my final result.