The issue with AIs reverse engineering code is that context is very important - in fact knowledge and understanding of the context is one of the few things humans can still bring to the table.
Unless every relevant fact about that context has been encoded in a recoverable way the system and tests, AIs can only do so much. And there are essentially no non-trivial systems where that's the case.
Absolutely you have no control over what others have written, but you also have no way to access their lost context, so you are no further ahead than an LLM in that situation. The available information is the same for you as any other system.
Maybe. If I know that the performance of this particular code path is going to be critical to the project's future success, sure. It's more common for something like that to be premature optimization though and the extra code is dead weight. I am not convinced by the idea that LLMs make that kind of dead weight much less undesirable.
An experienced developer will often have a good intuition about what might deserve attention in the future but isn't worth the effort now.
It's also useful for social reasons. Maybe the CTO wrote the original code and a junior developer working on the optimization thinks they know a better way but isn't sure about questioning the CTO's choice of algorithm. A comment saying it was arbitrary gives them permission.
If changing the algorithm is going to negatively affect the program then the CTO would have written tests to ensure that the important property is preserved. There is really no reason for the junior to be concerned as if he introduces an algorithm that is too slow, for example, then the tests aren't going to pass.
Yes, it is most definitely possible the CTO was a hack who didn't know how to build software and the junior was brought in to clean up his mess. However, in that case the information is simply lost. An LLM will not be able to recover it, but neither will a human.