The visual tests show that the results from closest point aren't completely expected. Some polygons' results are along the expected edge, but others are on the opposite side of the polygon. This appears related to where the query point is so perhaps this is a result of winding order or a negative value origin point?
Here's a gif showing the results from vtquery flipping as the point gets closer to the polygons on the lower portion of the image (this is not on a tile border, FYI):

My expectations:

These are Mapbox Streets tiles and I can recreate the issue at z15 with any buildings.
cc @flippmoke