If I have locale set to C, I can do this:
regression=# create database u8 encoding 'utf8';
CREATE DATABASE
regression=# create database l1 encoding 'latin1' template u8;
CREATE DATABASE
Had I had any actual utf8 data in u8, l1 would now contain
encoding-corrupt information. Given that we've tried to
clamp down on encoding violations in recent releases, I wonder
why this case is still allowed.
(In non-C locales, this will typically fail because the two
different encodings can't both match the locale. But I don't
believe it's our policy to enforce encoding validity only for
non-C locales.)
We should presumably let the encoding be changed when cloning
from template0, and probably it's reasonable to trust the user
if either source or destination DB encoding is SQL_ASCII.
In other cases I'm thinking it should fail.
regards, tom lane