We get fairly close here if we can get providers to DKIM-sign Autocrypt headers, like Posteo already does.
E-Mail has a lot of baggage though, and we don't expect to get everything done overnight. Legacy is important, and development resources in this area are realistically very limited. Still, we feel like we have a reasonable start.
> It also needs to be robust against malicious MTAs who could otherwise silently spoof senders and misrepresent their keys (double signing?).
I think the best we can do here is to make providers sign keys during transit, and increase attacker risk by allowing users to verify their communications at any point.
They won't accept an e-mail from a user ("submission") without STARTTLS, however. An MUA must negotiate the encrypted channel before even being allowed to authenticate. AFAIK, this is pretty much a universal requirement on the {IMAP4,POP3}-over-TLS ports (993,995/TCP) -- it definitely is on the mail servers I run.
> Don't accept non-STARTTLS traffic. Google doesn't, ...
is incorrect -- as you doubted:
> If that's true, I'm both impressed ...
Don't accept non-STARTTLS traffic. Google doesn't, you shouldn't have to either. STARTTLS is widely available, and simply needs to be configured. Encouraging everyone to configure STARTTLS correctly is the easiest win we'll have.
And I hope you don't take what I said to mean that IMAPS and SMTP over TLS are a proper message encryption system. I simply meant that HTTPS doesn't do anything right that IMAPS or SMTP over TLS don't also.
Clearly there needs to be a standard way of advertising, distributing, transitioning, and retiring keys for email addresses. It needs to be robust, in continuity with existing infrastructure (probably domainkeys signed), straightforward (maybe signed proof of sender keys in a header), and possible to delegate in a way which meets the requirements of mailing lists. It also needs to be robust against malicious MTAs who could otherwise silently spoof senders and misrepresent their keys (double signing?).
A system with these properties would also solve the problematic interactions between DMARC and mailing lists which prevent almost everyone from deploying an appropriate DMARC policy today. If somebody is working on this I'd like to know about it.