Hmm -- I don't think there's any flushing of buffers that could cause memory problems. However, there could be things related to uploads. If you can recall any other things of interest that are happening before memory is gobbled up, it'd be very helpful.

Splitting the sources into two is a great idea, but one that is hard to implement in practice. If you have four sources and try two of them, what happens if you succeeded a bit with them and then they went offline? Do you fallback to the other two? If so, and corruption is noticed after it's over, which sources do you retry? It would be much simpler to just do it the way it's supposed to be done -- verify as it's downloaded, but it's not possible to do that right now. It _will_ be done that way at some point, though.