Owen did some memory statistics and found out that the SPEC benchmark 447.dealII had a 13% reduced memory footprint with my patches. This is about what I expected, and a pretty awesome result. Despite of the algorithmic overhead that has been (knowingly) added the compilation did not really get slower, indeed some nightly testers show slightly improved cpu times. Again, this is in line with my observed slight speedup on kimwitu++ benchmark's compilation time.
My plans are to do some more API cleanup and then go after even more memory savings which I am already drafting in my fantasies... Whether they are feasible compile-time-wise is still an open question, but I am determined to find out.