The crucial assumption here is that all the 200 landmark points are in correspondence. Otherwise GP Regression would not work.
We have indeed run the non-rigid ICP algorithm to establish correspondence between the given fragment and the model to obtain this correspondence.
Once correspondence was established, we simply picked 200 points at random.
The posterior model we compute in tutorial 9 is the same as we compute in the non-rigid ICP. For the non-rigid ICP we use, however, only the mean whereas in tutorial 9 the main point is that we obtain the full distribution (another GP).