The other distinction is, if a client-side form validator detects a problem with the input, it tells the user so they can fix it, whereas if Apple's system detects a problem with your input, it potentially tells the police without giving you a chance to delete the maliciously-generated false positive files you've unwittingly received.
What I mean is that client side form validation is typically part of a cooperative process - I want to buy something at your store, you want to sell it to me, and smart client side validation can make that faster/easier if I typo something it can catch. It is (or at least should be) mostly aimed at helping the user; your parsing for security is (should be), as you note, on the server.
Scanning photos for [CSAM, thoughtcrime memes, poor taste, whatever] doesn't make the user's life easier, is not something anyone asked to be subjected to, and potentially can lead to a very negative outcome for them.
That's the distinction I was getting at, and yes, your second point is directly relevant there.
It is hardly difficult to draw a distinction between ensuring a field looks like an expected datatype and ML analysis guessing at photo content.
In fact, trying to construct an argument conflating the two pretty much immediately runs in to the fact that one is adversarial, so it only works if you studiously ignore intent.