With http-01, ownership has to be proven every time a new cert is issued.
> Such an attacker would likely have also stolen the TLS private key, but that only stays valid for 90 days
That doesn't have to be the case; the private key can be valid for as long as someone wants it to be, unrelated to the validation period of the cert that is issued. Yes, it does look like certbot generates a new keypair for every renewal, but in a world where we were putting a pubkey in a DNS record, the private key would certainly have a much longer validity, as otherwise there'd be no point to doing it this way in the first place.
If anything, as far as I'm concerned, DNS-based challenges are a stronger proof that I own the domain as it requires access to the nameserver. HTTP-based challenges just prove that I have access to a computer pointed to by a DNS record which is far easier to get wrong. This is why wildcard domains cannot be issued to HTTP challenges, just because I can serve a file on a subdomain doesn't mean I own the parent domain.
But I agree, using a key-based DNS challenge would be a new feature. It was discussed before but at the time LE devs didn't come to a consensus with how to move it forwards.
In addition, the new work mentioned in this letter is ARI, or ACME Renewal Info, which is not directly tied to any of the aforementioned renewal methods.
Although it does require an additional NS which is an additional maintenance overhead - that is preferable to leaving automated services sitting around with write access to important domains. At least until I can stop regularly writing DNS records.