The short answer is unfortunately 'NO'. I ended up removing all annotations from our figures since this was the only "quick fix" available. Rather annoying, but not critical for our product.
Our code is therefore currently not leaking memory but for safety I have added regular calls to the Matlab MEMORY function. If the Matlab memory usage exceeds a predefined maximum value the MCR executable will do a graceful exit and terminate (this seems to free the memory). A simple DOS batch script will then re-launch the MCR executable.
This is a really ugly solution but it's better than an unpredictable crash.
My hope is that Mathworks sooner or later provides a fix since they have confirmed the problem to exist in 2012a. If I (ever) find some time I would really like to do a more thorough analysis on 2012b and push Mathworks for a solution.
If you have more luck with 2012b please drop a note.