If I delete something from FaceBook, there's no way anyone (other than FB) can prove that I have said the things I deleted. With a signed append-only log, you'll have to disown your entire log and your keys. This is the problem.
B: You can't prove that I said X! Here's where I published private key used to sign that message.
A: Do you normally publish the private keys of your messages?
B: No
A: Okay, while I agree I do not have cryptographic proof that you said X, I think it's extremely likely that you said X and then later attempted to retract it by publishing the private key. So I'm going to go ahead and act as if you said X.
It also doesn't solve another part of the problem with append-only, accidental oversharing. Even if we take this 'publish the keys' process as actually working the data is still out there so if you publish info you shouldn't like locations or PII it's always out there.
That's true no matter what, purely by virtue of being shared with others.
I believe there are point-to-point protocols that publish every private key after the recipient receives the message, but I don't see how it generalizes to publish-subscribe.
I'm missing something..
That's why the key should be published as soon as possible. In a point-to-point protocol, it would be published as soon as the recipient confirms receipt of the message. That's why I'm not sure how it generalizes to group chat protocols.