See the discussion in CGAL-discuss:

• in Nabble: http://cgal-discuss.949826.n4.nabble.com/CGAL-3D-mesh-generation-prerequisite-condition-td4661693.html
• in the Sympa archives: https://sympa.inria.fr/sympa/arc/cgal-discuss/2016-03/msg00035.html

When there are no 1D features to protect, the initialization of the mesh from a polyhedral domain might not be good enough.

One possible solution, proposed by @afabri

Andreas proposed one way to ensure that the initialization points are correctly spread over the whole surface of the domain:
https://gist.github.com/afabri/1bbf46c6fd4f742bd982

Another possibility: check the result

There should be a way to:

• check if the output if good enough
  • Haussdorf distance?
  • ...actually, no need for a symmetrical distance, because the criteria of Mesh_3 already verify an approximation of the distance from c3t3 triangles to the boundary of the input domain.
• then possibly re-insert new points, and re-run the meshing process
  • the strategy for new points could be: if in the input surface there are vertices that are far from the c3t3 (or just from the vertices of the c3t3), then insert recursively the farthest points