The tricky bit was implementing Delaunay triangulation in what is, I think, the second-most horrifyingly naive way possible. bhudson, you probably want to avert your eyes. I'm not directly using the circumcircle definition of Delaunay-ness, but the thing I did is still definitely O(n3). I believe optimal (based on cursory Wikipedia reading) is O(n lg n), so that's pretty obscene indeed.
And then I filter out some nodes that are too near each other, and those that are too collinear with each other, and assign edge-orientations in a recursive way that guarantees the whole graph is strongly-connected, and there you go.