I wonder about the utility of a facility like this when it has such
weak guarantees.  Can you give an example of a situation when this
would be useful without having confidence that the arity information
is accurate at least for "most" procedures?

It also seems that including a means by which the programmer can
declare arity information for procedures (perhaps including system
procedures) could significantly improve the accuracy (and therefore
usefulness) of arity information, in particular about the arities of
procedures produced by procedure combinators.  In the presence of
mutation and procedures being comparable by eq?, such a mechanism
could be implemented with a (weak) eq? hash table on procedures.
(For what it's worth, declaring arity information for procedures is
semantically different from restricting the arities of procedures,
and may therefore enjoy broader use).

Best,
~Alexey