I think this idea is a nonstarter, TLS or not. We're generally moving in the direction of never letting the server see cleartext passwords. It's already possible to configure libpq to refuse such requests (see require_auth parameter), although that hasn't been made the default.
ALTER ROLE xxx WITH PASSWORD accepts hashed values, so a client with the SCRAM-SHA algorithm could:
1. remember the password that was just used to log in,
2. generate the new hash,
3. send that as an ALTER ROLE statement.
Anything which shows up in the logs would be no different than when someone types ALTER ROLE ... WITH PASSWORD from the psql prompt.
