And a side comment that by a common-sense definition, a list of
points that's not coplanar is not a polyhedron face.
Most non-trivial sets of more than 3 points are not planar due to floating point approximations. And some measure is necessary to cope with the different number representations of OpenSCAD and CGAL. As far as I know, OpenSCAD triangulates every polyhedron facet with more that 3 vertices precisely to avoid non-planarity issues. It can be argued that its triangulation is not "optimal" in some sense but we cannot blame it for doing this.
Polyhedron is a very flexible primitive. It is possible to define a non-manifold and/or singular surfaces even with borders or multiple shells with polyhedron. If degenerated faces are avoided, you can render it nevertheless while no boolean operations is performed. That may be a hell or a heaven depending on what you are doing or want to do.
What Uwe's workaround brought up is that the hull operator may "heal" (most) defective polyhedron generating a valid solid with a manifold shell: the convex hull of the polyhedron face vertices. That is nice, because mathematically the hull operation does not require any structure linking the points.
I don't think there is anything wrong with the current processing of polyhedron. But...
Consider the following:
e = 5;
p = [ [0,0, 0],[10, 0, 0],[0,10, 0],[0,0, 10],
faces1 = [[0,1,2],
faces2 = [[4,6,5],
This polyhedron is really the union of a pair of tetrahedra provided that the parameter e is positive. It can be boolean operated with another primitive and renders fine. However, when I set the parameter e to 1e-6 and union it with a cube, my 01.06 version running on Windows crashes. Worst than that, setting e to 1e-5 I am able to render it unioned with a cube and generate a STL file. Although the file is not defective (it represents exactly the two tetrahedra with a small gap of about 1e-05) OpenSCAD crashes when I try to import it.
Summarizing, my points are:
a) it is fine to define polyhedra that cannot be boolean operated because they are not manifold;
b) it is fine that those polyhedra produce STL files that are not manifold provided a warning is issued;
c) it is not fine that a STL file representing a manifold exported by OpenSCAD cannot be imported by OpenSCAD.