Why should it work?
For example, we can try this:
XYZ1 = dec2bin(0:7) - '0'
So we have the 8 vertices of a unit cube. There are subsets of these points that are coplanar. Of course, delaunayTriangulation will have no problem with the set.
T1 = delaunayTriangulation(XYZ1)
Howver, suppose we try a simple set of points that lie entirely in a plane?
XYZ2 = randn(8,2)*randn(2,3)
These points all lie in a single plane. We can see that is true, by looking at the results of the singular value decomposition. It will have one zero element in S.
I should rotate it around to show the points lie exactly in a plane, but I'm too lazy now.
T2 = delaunayTriangulation(XYZ2)
See that delaunayTriangulation survives, although off those simplexes will have zero volume.
However, some of the older tools will fail.
Effectively, the problem probably arises from the older codes, which were dependent on qhull libraries. It seems that tools like delaunayTriangulation survive. Note that a common solution in those older codes was to use a joggle, so a tiny additive random perturbation, that would result in the points no longer being coplanar.