The problem, I think, is that he wants a low level tool for generating disjoint types (not satisfied by define-record-type, which is not required to be disjoint from vectors or conses or what have you). This would give you the ability to have enums, ML style discriminated unions, simple records, etc. without necessarily requiring that each and every one of them be in the small language.

That said from a practical point of view some sort of record type is long overdue.