Bingo, Dan's got it. I remember I had seen this problem before, and so it seems that the empty string gets interpreted as "0" for the integer type. Don, we should report this as a bug, it really should throw an error, which would make our porting effort easier here.
I don't think the solution is to change empty_string_p, because the problem is with the data layer: it's wrong to keep a 0 in the database where you should have a NULL. It means more work to be done on our end, but it's better to fix the data entry than patch the data selection.
