Today's release doesn't have the database branch merged of course, but it still has a significant amount of changes.
Developed a test case for the sqlite problem, that reliably reproduces it, and sent it to the sqlite mailing list. It seems that under heavy write load, when a new connection is made to the database, SELECT can fail for a little while. Once one SELECT succeeds, that database connection becomes solid, and won't fail any more (apparently). This makes me think there might be some connection initialization steps that don't end up finishing before the SELECT goes through in this situation. I should be able to work around this problem by probing new connections for stability, and probably will have to, since it'll be years before any bug fixed sqlite is available everywhere.
I also noticed that current git-annex incremental parallel fsck doesn't really parallelize well; eg the processes do duplicate work. So, the database branch is not really a regression in this area.