So, the superficial argument seems to be that this is just an editorial confusion between direct vs. base-class-recursive non-static data members. Or it could be considered an oversight when default member initializers were made to work with aggregates. @zygoloid, this seems borderline editorial.

it could be considered an oversight when default member initializers were made to work with aggregates

I think default member initializers are not the key here, class B { int i; }; is also fine (updated the PR message). It looks like an oversight when aggregates were allowed to have base classes (http://wg21.link/p0017r1). Note that the definition of aggregate elements says "direct non-static data members". (Also, the proposal for designated initializers (http://wg21.link/p0329r4) speaks about "direct non-static data members".)
Also, the note in [dcl.init.aggr]/1 — "Aggregate initialization does not allow accessing protected and private base class' members" — suggests that the item does mean only direct members.

This seems editorial to me. When we allowed aggregates to have base classes, the design intent was that base classes are just regular aggregate elements like non-static data members, which I think makes it reasonably clear that this wording should be interpreted as applying only to the direct members.

So, let's apply this.