The mentioned method has no specification and it is very hard to find out how
it is supposed to work.
We have to wait for a free software application that makes use of the class to
get the implementation done.
Additionally it is known that the current implementation will fail with right-to-left written text.
I think the supposed behaviour of this method is not too hard to infer: this should be used mainly for caret movement via the cursor keys, so you can think of it as a means of finding the next (document) position in the specified direction. For east-west movement this should normally be quite trivial, for north-south movement this is slightly harder but not too hard too.
I cannot see how our current impl is wrong. I set the severity to enhancement and leave the bug open for now.