I've noticed some odd behavior in the DM42's reported free memory as well. I'm guessing (i.e. I haven't read the source) that Free42 is just reporting the amount of memory that OS tells it is available, in other words however much it could malloc for its own use. You'll notice that editing a program and adding a few steps doesn't immediately change this value, so Free42 is presumably allocating memory in chunks, then using those chunks to store calculator data. The amount free will probably vary depending on what the DMCP OS has allocated for its own use, any garbage collection that happens either in the OS or in Free42, memory fragmentation in Free42, etc. Then this is all compounded by the fact that Free42 consumes memory faster than a real 42S, both because of the data registers being larger and higher precision, and program steps not being packed as tightly (a program that is reported as being 100 bytes is actually occupying about twice that, for instance).
In short, it's really hard to tell exactly how much free memory you've got left on the DM42.
The same would be observed in Free42 running on other platforms, but since those platforms generally report several gigabytes of free memory, it's of little consequence.
I haven't done any empirical tests, but it "feels" like I can fit more in my upgraded 32 KB 42S than I can in the DM42, based on how quickly the reported free memory decreases.
EDIT: This might be the other thread you were thinking of, where I was discussing memory usage and allocation with Thomas.