Machining feature recognition is a still an active and quite narrow research field so I'd be surprised if you get any out-of-the-box solution here. The best you can do IMO is to keep trying out solutions from the literature and see how they work for you.

You could indeed get a pointcloud by sampling your original mesh, compute local descriptors and use them to train a classifier. But there is no guarantee that the local geometry around a point is enough to recognize a machining feature.

The solution described in the paper Freeform Machining Features: New Concepts and Classification uses a mesh as input geometry and combines differential geometry and graph theory to classify machining features. Maybe worth a try?