Put simply, the problem is this: in order for Epenthesis to only be triggered by the prospect of complete identity of adjacent consonants, voicing and all, this rule must be able to consider what would be the output of another rule; namely, Devoicing. (Another way to look at it is that Epenthesis must be able to consider what would otherwise be the output of the entire derivation; these two alternatives cannot be distinguished in this case, I think.) This is very different from the far more limited sense of ‘peeking’ proposed by Hill (1970), who writes (p. 534, emphasis added):
Cupeño […] offers an example of an unusual type of rule which, in order to be properly stated, must apparently have access to its own output; knowledge of the phonological input to the rule is not sufficient to state the correct generalization. I propose the term ‘peeking’ for this kind of behavior.
Hill’s example is just one of many from the early 70s that exposed the limitations of the assumption that phonological derivations are strictly Markovian; that is, that “a rule applies to a form if an only if the form fits the structural analysis of the rule at the point in the derivation at which the rule is applicable” (Kiparsky 1973:57). Kiparsky’s paper is of course a serious (and seriously interesting) attempt to find “general conditions” on “looking back” and (as with peeking) “looking forward” in derivations; that is, to limit the power of such mechanisms, which Kiparsky fully acknowledges are necessary though preferably not unlimited in power.
[ I should add here that Kiparsky (1973:83) notes that Hill’s example is “particularly interesting”, but that it is one of several otherwise uncited “scattered examples of other types of global-looking phenomena in the recent phonological literature [which he] will not try to make any guesses about at this time”. I may just be ignorant of the relevant literature, but I don’t know of any specific work building on Kiparsky (1973) that specifically limits (or aims to limit) the power of “looking forward” such that Hill’s type of example (and perhaps a well-defined set of others like it) are possible. If you know of such work, please let me know. ]
I gave this matter quite a bit of thought as I prepared an early version of this paper for presentation at SWOT 6 in 2003. The best derivational solution I could come up with is sketched in my SWOT handout (see section 3, starting on p. 3). What I concluded was that, even given some interaction between rules and constraints (as is generally assumed in derivational phonology), the way in which they must interact is rather unusual; certain false steps must be taken in the derivation in order to figure out that one or the other of the constraints is violated, at which point the derivation must ‘backtrack’ to undo those false steps. The OT analysis must also take false derivations into account, but that’s what you always do in OT: you compare various output candidates, all but one of which are nonoptimal (= false), in order to figure out which is the optimal output.