The issue was fixed in 8.5.2FP2 and 8.5.3 onwards. More details here.
http://www-10.lotus.com/ldd/fixlist.nsf/Public/E461469D3BEA868E8525783000736F00?OpenDocument
I'm not aware of a regression.
If you break it down to multiple subsets of the documents does it work? If not, is there a particular document? And if so is any Base64 data correct/too large?