What you're suggesting is going to lead to situations where the user
sets a non-default collation on every field in every table in the
database and depending on the query they will sometimes get the default
collation anyway.
Not that I know a lot about this, but I agree there should be some kind of bubbling up here. ISTM that you could think of this as replacing the text type (and maybe others) by a collection of closely related types, and operators like >(text, text) become parametrically polymorphic. I am not entirely convinced that there won't be corner cases when this implicit polymorphism will get it wrong, but if it does we can apply a suitably sized band-aid. I agree with Martijn's analysis that it will get it right a lot more often; and consequently avoid the need for a lot of manual fiddling.
...Robert