The answer depends on which definition of "phoneme" you use. Under the classical taxonomic definition, where you analyze actual sounds into a more abstract system, two sounds are allophones if their surface distribution is complementary, and if their surface distribution is contrastive, they are distinct phonemes. The concept of "conditioning environment" is a secondary concept, related to the rules which relate surface sounds to a system of phonemes. Under the classical view, if the conditioning factor is later changed (especially deleted), then the relationship is no longer allophonic, because the surface distribution becomes contrastive. In OIr, not every word-final rhotic after [i] is palatal, therefore plain and palatal r are separate phonemes (although the palatal has a restricted distribution).
The generative reinterpretation of "phoneme" is "any feature matrix that can be present in the underlying representation". Under that view, if every instance of a sound X can be derived from some other sound by a rule, then X is not a phoneme. So if the underlying form of "man" is /wiri/ and there are appropriate rules to derive palatalized r, then palatalized r is not a phoneme. Surface loss of the conditioning environment does not change the phoneme status of the sound. What does change its status is whether the rules and underlying forms required to deny a sound phonemic status are still learnable. For instance, when the final vowel of "clothe" was actually entirely lost from the grammar, the rule allophonically deriving [ð] from /θ/ became untenable, and /ð/ became a phoneme (there were other factors converging on that result, it was not just about vowel loss).