If the social contract says "delete things when you see a delete message", then that's useful in a federated environment.

Only bad actors will disobey and they will have to modify their software in order to do so. This doesn't provide absolute protection against bad actors, but since most bad actors don't own a time machine, it reduces the scope of the harm they can enact.

Consider the adversary "angry ex-boyfriend." Let's assume he wasn't always angry and isn't a sociopath. By the time he has become angry, the sensitive posts have already been deleted. This makes a difference in real life to real people.

There is indeed a user-interface concern, to not over-promise to the end-user. But that doesn't justify leaving such a useful thing out at the protocol level.