I've been told that "the word-final affricate dź in Polish should be devoiced to /t͡ɕ/". What are the linguistic precedents for this, and why is this devoicing not evident in the Wikipedia pronunciation?
To be precise, it is /wudʑ/ and in most contexts it will be pronounced [wut͡ɕ]. Phonologically there is phoneme /dʑ/ because when you decline the word, there is a vowel that follows, the phoneme stays sonorous, e.g. "Łodzi" [wodʑi]. But otherwise, the sonority is lost/neutralised at the end of the word, so unless the following word starts with voiced consonant, the final segment will stay devoiced.
For the recording in question, this may be dialectal or just an example of the word taken from the middle of an utterance.
There are many explanations for the properties of a given random sound file on the internet. Since Wiki provides virtually no metadata on recordings, one can only speculate as to the circumstances of a recording. That token certainly doesn't tell you anything about the structure of the language. One step in the direction of making the question be about the language and not the sound file is to get other examples. Wiki also provides this token which BTW is linked in the Polish page, where the stop is clearly devoiced. All of the Forvo tokens have devoicing. Other Wiki entries with potential and actual devoicing are Tarnów, Grudziądz, Stargard and Bełchatów. So by overwhelming majority rule of evidence, Polish does have final devoicing.
To get the truth, you need more-controlled data, and you need to control for well-known artifacts. The most important is hyperarticulation, especially a problem with intelligent and multilingual speakers (you can interpret "and" as "or"), particularly a problem with literary languages. That is, the grammar of Polish does not prevent speakers of Polish from articulating things not generated by the grammar of Polish. People often have normative ideologies that control how they behave in guarded moments (like, making a recording). The other variable is phonetic context. In the offending file, you will notice that the right edge is prematurely clipped, which suggests that the recording was extracted from a longer utterance – it was not the pronunciation of the word in isolation, it was in a phrasal context. Since Polish has voicing assimilation at the phrase level, it is quite plausible that the token came from a voicing context. Thus you would need to get actual citation forms, not phrasal forms that were edited down.
The best would be just to hear the native pronounciation: https://de.forvo.com/search/lodz/pl/