It would be crazy overkill for this type of problem, but I think the functionality of the generalized Hough transform for object detection at least encompasses what you want, even though it's not quite meant for this purpose.
Given an arbitrary shape created from a set of points, and another arbitrary set of points, it tries to find the shape in the set of the points even though it's been rotated, scaled, and translated. You might be able to take out the scaling and translation and get what you want.
Basically what it would come down to is brute forcing possible rotation points to see which one fit the second set of points best.
