This is a nice idea.  If I may be so bold, you might also consider using my HAMT library, published with SRFI 146.  I haven't thought through the implementation, but my intuition is that you could then provide a persistent version of this data structure, too.