Grahame has asserted that filters should not be allowed for required bindings to 'code' elements and the tooling currently enforces that.  However, it's not a rule in the spec.  We should either agree that it is a rule or we should update the spec.

Personally, I don't see any reason why it should be a rule.  The expansion can be safely calculated at publication time and is frozen, which means downstream implementers don't actually have any extra work to do...