This is a fix of the bug mentioned in #1333. When projecting to Mollweide, for latitudes above about 89.14 (or below -89.14), all points will be mapped to the pole. This causes potentially serious errors for things like contourf. My solution is to modify the boundary so that the pole is on the outside and no points on the boundary are mapped to the pole.
From what I understand about how projecting segments works, when a line segment is projected so that it crosses the boundary, cartopy will try and find a point on the original line which, when projected, will lie on the boundary. This means that so long as the boundary is outside of the problematic area, all projected line segments touching the boundary will be significantly more correct than what they would have been had they instead continued to the pole. This ought to prevent many of the problems with this projection.