There's a broader issue here: the lookup for a dependent qualified name is erroneously said to take place from its definition, full stop ([temp.res.general]/1). Obviously in simple examples like

template<class T>
struct wrap {using type=T::type;};

struct A {using type=int;};

wrap<A>::type x;  // OK

the lookup for type in the definition of wrap<A> must take place from someplace after the definition—presumably it uses the point of instantiation, which should be sufficient for reachable even though we need the whole instantiation context for ADL.

Once we (I) fix that, we should come back to this to make sure it says what we want. (I know not everyone was happy being able to use lookup into base classes in other base-specifiers anyway.)

Actually, this issue is accompanied by #4776. In that issue, I claim the definition of a class introduces a class scope rather than merely declaration does, which seems can resolve we do not lookup into base classes in other base-specifiers. I also tried to feedback this issue to core language issue, while his opinion is

The use of "declaration" there was obviously intentional, as indicated by the "(if any)" notation - the declaration is acknowledged as possibly not having a member-specification. In any event, it seems harmless to introduce the class scope with an elaborated-type-specifier, even if it will be empty. I think that it simplifies the description of qualified lookup - you look in the scope denoted by the nested-name-specifier, which will simply be empty if the class type is incomplete. You don't have to specify what happens if the nested-name-specifier, although declared, is not associated with a scope, because there's always an associated scope.

It appears to me that such an opinion is stressing this issue #4776 (comment) I have asked.