I disagree that this is an exception, because the database is full of simple, 1-way closed polygons that have boundary tagging but are actually an inner or outer member of a boundary relation. In such cases, it is ambiguous as to whether that polygon represents its own unique boundary, or is mis-tagged as duplicating the boundary tagging of its parent relation. This is especially problematic in cases where the simple closed way contains a name that is unique from the parent boundary relation, e.g. “<name of city> Boundary”.
That is why it’s necessary for boundaries to be a relation in all cases for them to be well-formed and understood unambiguously by data consumers, and why validators (including the one I developed for US boundaries) report this as a problem.