So from the editing process the AI thinks its just a file. It likely doesnt know it cant do it until it tries, as is with any file.
Unless you can share the attempts the AI does I can only guess what the AI tries and how it tries it. There are technically several ways to edit files that the AI models in Agent uses. Also each AI model has slightly different behavior patterns due to its training, but also dependent on the prompt to a part.
As a user I understand your point of view. Please note that some users prefer that AI doesnt touch those files. We can assume that Cursor has to make sure those features work and dont break so most users can use them without having to constantly check why it isnt working anymore.
For me, I dont change the rules that often once they are working well.
If you are not just changing the rules but trying to store progress or similar info I would suggest following post by another user, see top there. Perhaps also as a workflow suggestion.