For what I understand, this would fall under the category of partial shape retrieval/matching problems.

A classification was proposed already some time ago in Tangelder, J. W., & Veltkamp, R. C. (2004). A survey of content based 3D shape retrieval methods. Proceedings Shape Modeling Applications, 2004., 145-156. Basically the resolution methods can fall into a few distinct categories such as: feature based, graph based and geometry based. Only local features and graphs are adapted to partial matching according to this article.

For potential experimentations using Python:

• Local features can be computed using for example PyMeshLab (it can compute different geometric information like curvature, geodesic distance, or local vertex density) and then somehow compared with existing meshes. This comparison step is the hard part in my opinion and I don't think there is a good answer.
• Meshes are already graphs (vertices connected by edges, forming faces) so you could indeed compare graphs, using for example the VF2 algorithm of the NetworkX library. However, note that the "subgraph isomorphism problem is NP-complete" as said in the doc so concretely it doesn't scale well. You could also try to segment your mesh into patches (planar, cylindrical, etc.) and make a graph out of it, but same limitation.

You could also use a matching/registration algorithm such as the ICP provided for example by Open3D. This is however very sensitive to the initial alignment/position of the meshes.

Another solution would be to use deep learning and segmentation. If you have a list of examples that is big enough (and a lot of time to correctly label your data), you could try sampling your mesh and using the PointNet model using for example Keras. This would result in a list of labels associated your points that are themselves associated to a particular class like "pin" or "hole".

But unfortunately for you, I don't think this is a "pretty standard task" :(