[temp.res.general]/6.4 is not talking about substitution failure, it's
talking about satisfaction checks of atomic constraints being
ill-formed (a situation that arises when substitution succeeds but does
not produce a constant expression of type bool). If a satisfaction
check of A is performed, then the program is just ill-formed
(diagnostic required).

Since CWG2461 there's no equivalent of [expr.prim.req.general]/5 for
atomic constraints. (Maybe there should be; I'm not entirely convinced
the removal was intentional.)