The documentation stresses its raw speed compared to some other Boolean packages out there. While that is well and good, how about reliability? The few examples pictures there seemed to highlight speed - the actual examples in these cases did not visually appear to be “difficult” Boolean operations.
There’s also a few other angles, of which robustness is only one. Some/many CSG libraries don’t handle the interpolation of attributes other than position, leaving it up to you to rebuild uv mappings, and recomputing normals, and materials.
And generally, CSG is a process that doesn’t have many guarantees, and also has endless edge cases. Literally flipping the orientation of a single triangle in any of the input meshes can destroy the whole operation and yield garbage/infinities. The input meshes generally have to be “manifold” and “watertight” to get reliable results.
If it has to be fast, and support material boundaries/normals/uvs, etc… then three-bvh-csg is the best choice imo. If robustness is a concern, then that manifold library might be good since it’s modelled after OpenSCAD. Not sure if it handle materials/uv/normal interpolation.