Yes that would probably work. There could be some essential features of weapon parts that an algorithm might still be able to learn, and a printer could also keep track of previously printed parts for the classification. I think its unlikely that there are essential features of gun parts that are specific to gun parts so there would probably be a lot of false positives.

Isleepinahammocks idea would probably work. But rotation and translation would not. Thats something you can easily take care of in your training data, by reusing the same training data in multiple random positions and random angles.

Yes that’s probably how you would do this. Get a bunch of data of gcode of 3d printed gun parts and not-gun parts, for different slicers and printers. Then train some transformer as a classifier. Based on how good object recognition is, i would say its possible that you would get reasonably good accuracy and precision. And because you are scanning for code the architecture will likely be similar to an llm.