After borrowing a friend's MacBook and testing, I can confirm this issue.
It is not an issue specific to version 2.3.3, and actually doesn't even seem to be a bug in the code at all, but for some reason the OS X binary for version 2.3.3 produces 1/3 wrong hashes for algorithm scrypt. Binaries for all other platforms are not affected. The problem appears to be caused by a bug in the particular compiler used to produce the binary (Apple's version of Clang for OS X Lion), and is only triggered at -O2 or higher optimization levels. The issue does not appear if using gcc on OS X, nor does it appear with the latest version of Clang on Linux.
I have now updated the OS X package for version 2.3.3 with a binary compiled at -O1. It would be nice if you could confirm that it works.
Seems to be working fine now with updated version 2.3.3 you posted overnight. Thanks!
I know CPU mining is essentially dead, but it's nice to know you still maintain this. Cheers!