Complex DR definition works as at-rule

[seven-phases-max]

A variable definition with a value that ends with an unnamed ruleset, e.g.

@var: x {y: z};

is interpreted as an unknown at-rule, thus resulting in:

@var : x {
  y: z;
}

This is most likely a result of #2783. (Before #2783, such code resulted in a error, i.e. it's not a regression but more like an overlooked side-effect).
See also related #2270 (comment).

[seven-phases-max]

P.S. Technically, there is an at-rule that has such syntax, i.e.:

@page :first {...}

So the new behavior is not that wrong (not counting that whitespaces in @page are significant and in @var they are not), but if kept, it anticipatorily kills possibilities mentioned at #2270 (comment).

[rjgotten]

Pseudo-class selectors don't allow whitespace though, so you could disambiguate on that:
i.e. @var: x {y: z}; vs @directive :pseudo {y: z};

Following that, the exact example given in your opening comment should not be interpreted as an at-directive, but as a variable.

(Moreover, @page is kind of a special snowflake as well.)