CREATE TABLE foo ( a CHARACTER VARYING CHARACTER SET UTF8, b CHARACTER VARYING CHARACTER SET LATIN1 )
and so on, and write character literals like
_UTF8'Hello, world!' and _LATIN1'Hello, world!'
and have those columns and data types independently contain what they can contain, without constraints imposed by one overall database encoding.
Obviously, we're far from being able to do that. But should it become desirable to get closer, would it be worthwhile to also try to follow how the standard would have it look?
Clearly, part of the job would involve making the wire protocol able to transmit binary values and identify their encodings.
I would go in the other direction (note: I’m ignoring all backward compatibility considerations related to the current design of Postgres).
Always store only UTF-8 in the database, and send only UTF-8 on the wire protocol. If we still want to have a concept of "client encoding", have the client libpq take care of translating the bytes between the bytes used by the caller and the bytes sent on the wire.
Note that you could still define columns as you say, but the character set specification would effectively act simply as a CHECK constraint on the characters allowed, essentially CHECK (column_name ~ '^[...all characters in encoding...]$*'). We don't allow different on-disk representations of dates or other data types; except when we really need to, and then we have multiple data types (e.g. int vs. float) rather than different ways of storing the same datatype.
What about characters not in UTF-8? If a character is important enough for us to worry about in Postgres, it’s important enough to get a U+ number from the Unicode Consortium, which automatically puts it in UTF-8. In the modern context, "plain text" mean "UTF-8 encoded text", as far as I'm concerned.