Travelling salesman with a directional constraint
Judging by the documentation on pytsp, the distance matrix doesn't have to be symmetric. This means that you could modify the L2 norm to incorporate information on a preferred direction into that matrix. Say you have a preferred direction for some pairs of points (i,j), then for each of these point you could divide dists[i,j] by (1+a) and multiply dists[j,i] by (1+a) to make that direction more favourable. This means that if your algorithm is sure to find the global optimum, you can force it to satisfy your preferred direction by taking a is sufficiently large.
Also, I'm not sure it's impossible to have closed loops in a solution where the distance matrix is taken from 3D data. It seems to me that the 'no closed loops' is a result (of the triangle inequality) specific to 2D.