The irreversibility of merger is an empirical principle, not a conceptual impossibility. In the right situation it could occur. A merger might be reversible under extreme conditions of dialect mixture, where speakers have sufficient exposure to (that is, nearly constant interaction with) speakers of a dialect where the merger did not take place, and further where the first speakers have strong motivation to learn the sound system of that dialect. This exposure must occur at a time during language acquisition when speakers are learning phonological categories and the lexical distributions of those categories. Further, the speakers of the unmerged dialect must not accommodate to the merged speakers by neutralizing the distinction in production, as was shown to occur under similar conditions of dialect contact by Herold (1990), in a study of the progress of the low-back merger in Pennsylvania. These conditions are sufficiently rare that they might well never be satisfied.
Mergers have always been found to expand at the expense of distinctions. This fact is the foundation of the comparative method of historical linguistics: if two related languages are examined and a distinction is found in one that does not occur in the other, then the proto-language from which they descend is inferred to have had that distinction.
With considerable effort as well as constant exposure, isolated speakers, immersed among speakers of another dialect, may be able to acquire a phonetic distinction and its lexical distribution which was merged in his or her native dialect.B.3 No law of reason or nature says that merger is irreversible. However, it is a far-fetched possibility that an entire community of speakers should have this much exposure and motivation, at the right age, for this kind of sound change to occur. If an entire community were surrounded by speakers of another dialect, they wouldn't be surrounded, because there would be as many of the members as there are speakers of the other dialect. Further, in this kind of situation, the powerful forces of covert prestige are likely to exert themselves to prevent assimilation to the other dialect. Finally, imitation of other dialects can be ``good enough'' to pass as a successful imitation when only a relatively few phonetic targets are imitated; an imitation can well succeed without being a fully accurate one.B.4 So if the goal is successful imitation of a target dialect, it can be attained without going to the extreme of acquiring a new sub-lexicon.
In summary, there is no clear evidence in favor of, and considerable reason to doubt, the reversibility of mergers. Merger is irreversible.