In general every time we look for some type of bad memory usage in Gecko, we find it somewhere. One thing we don't do a good job of tracking right now is memory allocated by mmap that doesn't live on the heap and doesn't live in JS. I don't know how much such memory there is, but I expect it's not zero per the rule above. I have a few orthogonal ideas of how to move forward here; I'll file them as blocking this bug.

Good idea! But note that once we've identified the dark dark matter, it'll then just be dark matter, albeit of kind "mapped-unclassifed" rather than "heap-unclassified" :)