Neat find. Is it possible that they're just stashing away benchmark results in some sort of large, inefficient data structure? We should try to get a break down of the mark time and see how much XPConnect is contributing, since I assume that's the sort of leak you're thinking of?
There could be all sorts of stashing involved, yes. Worth looking at the actual code too, perhaps.
Gregor, would you have time to look into this?
Downgrading this bug -- I don't think we're going to take any concrete action here unless we have a reason to think that Dromaeo itself is not to blame here.