This depends on the system. Mechatronics engineers spend a lot of time learning and figuring out how to make systems that can withstand edge cases like an incorrect screw size. There's whole engineering/mechatronics disciplines around automation reliability (though much of it involves lots and lots of sensors and cameras, haha).

The way they test these things is by intentionally throwing bad parts into the mix at various stages of their automation. Something like a screw being too big/small is a trivial matter that won't make it through a system or facility designed by professionals.

The real problem that really throws a wrench into every mechatronics engineer's carefully-planned automated masterworks is people doing things like throwing wrenches into their carefully-planned automated masterworks 😁

Why the flying fuck do you think I said, "non-general AI"? Even a well written algorithm could handle things coming in not in perfect shape, yet everyone pretends "non-general AI" means, "execute instructions repeatedly without any input what so ever."

Use your brain. Even basic dumb algorithms that can run on an Arduino can respond to input. Machine learning can easily respond to dynamic input, so stop failing to imagine the most basic of basic things I say.