The oldest form of Sanskrit has a class of expressions that are in some respects like asyndetically coordinated syntactic phrases, in other respects like single compound words. I propose to resolve the conflicting evidence by drawing on prosodic phonology, stratal optimality theory, and the lexicalist approach to morphological blocking. I then present an account of the semantic properties and the historical development of these expressions. The analysis points to a solution to the theoretical problem of nonmonotonic trajectories in diachrony, a challenge for causal theories of change that claim that analogical processes are simplifying or regularizing. The idea is that optimization of such a highly structured object as a language does not proceed monotonically, but via a sequence of local optima.