We currently prioritize kSoftClose/kArchive buckets over the kNone bucket and return them in a vector in such order.

We can further improve it with prioritizing kSoftClose candidates over kArchive, since we can easily reopen the kArchive ones whereas we're about to forget about kSoftClose (ignoring query-based reopening), so this is our last chance to fill them up. We could also prioritize buckets with less data in them, to improve fill ratios, etc.