Well, I was cleaning out the code, and it turned out to be impossible to maintain clusters without putting all the dirt back in. They were causing a steady stream of UI bugs and I could see no way around that.
So the next cluster implementation will suffer from none of those problems, but it's going to take a while to write it.