Ask quality to examine the identified bezel and locate any actual rejected mark. I'm no appearance inspector, but I wouldn't ask software to learn those labels while nobody can point to the supposed scratch on the part.
Keep that accepted-but-rubbed example. It sounds more useful than only keeping perfect parts. You'll still need confirmed rejects, and the camera comparison has to show their real features under the presentation you intend to use.
Compare the original captures before changing equipment. Were those two views made with fixed lighting, or did someone move the lamp while turning the bezel?
Zara, Anika's question is worth taking to tooling with the actual holder. A repeated image comparison needs to describe that movement, not just call every different picture a lighting problem.
Use those same identified parts when the holder returns. Otherwise a fresh sample pile will make the before-and-after comparison harder than it needs to be.