I observed that a lot of admin_level boundaries (2,3,4,5 and sometimes 6,7,8,9,10,11) are not useable because they are not completely contained in the tile data. The polygons are closed automatically but this is only a good guess in which direction they have to be closed. The probability is quite high that they are closed in the wrong direction. I think this is a fundamental issue with splitting and areas. Martin's suggestion about extracting boundaries and using them on the side makes sense, but I would say that the issue is about all large areas. I wonder if it's possible to evaluate areas (and things like boundaries which are semantically areas even if they are technically closed linear features) for overlaping with tiles, and to then replace the area by multiple areas projected onto tile boundaries.