This essay began to answer a straightforward question, one that appeared to be primarily about mere facts. But I wanted to provide a good answer to AMRs worthy question about the serious implications of incomplete neutralization rules. There are two main conclusions to be drawn from this discussion.
First, the lack of complete neutralization of syllable-final voicing in German is a real phenomenon. It has been observed many times. There appear to be several `sounds' here produced and perceived with only a moderate degree of distinctiveness. But whether they are the same or different depends on how you ask the question. The American English flapping situation is another similar example. There are others as well, such as the near neutralization of pairs like prints-prince in American English [Fourakis and Port, 1986]. As for other proposed cases in Polish and Catalan, etc, I am not in a position to make bold claims. Nor am I certain, in a dialect exhibiting an incomplete neutralization, that all speakers will necessarily show the effect. Speaker idiosyncrasy seems a possibility here. After all, who is likely to notice which way a particular speaker implements the `neutralization'?
Secondly, I agree completely with AMR that this effect is similar to other instances of the `near-merger' of a contrast (Labov and Steiner, 1972) and the gradual reduction of contrasts under `weakening' and `fast speech' etc. But more sweepingly, it seems that 50 years of experimental phonetics research on speech production, speech motor control, speech perception and descriptive phonetics have shown repeatedly that human speech sounds tend to distribute themselves rather smoothly over a wide range of variables. The incomplete neutralization result is just one particularly vivid and carefully studied example in a great mass of data supporting continuous phonetics over discrete phonetics.
AMR is right to be troubled about incomplete neutralization since it undermines a critical assumption about phonetics that is taken to be `gospel' by many working phonologists. It seems to me (as it does for many phoneticians and `lab phonologists') that, by trusting segmental transcriptions, phonologists run the risk, quite frankly, of building their work on sand. Such work tends to completely ignore everything about time except for what can be expressed in terms of the serial order of symbols. And they tend to ignore or represent only clumsily the many graded effects like `weakening', speaking rate and `articulatory laziness'. By ignoring continuous-time effects, phonologists run the risk of developing theories of the wrong phenomena and of overlooking important language-specific phenomena. Languages do exhibit some discrete `sound objects' which cry out for description and explanation: `distinctive features', `phonemes', `stress levels', `natural classes' (like obstruents, vowels, nasals, etc) and much more. But these are all phonological objects - linguistic objects - and it is linguistics that must find a way to explain them. Alas, phonetics cannot.
One finally remark on the practice of phonetic transcription: Despite my insistence that no alphabet for phonetics can be completely relied upon, I continue to teach the IPA alphabet to linguistics students. It is useful for very many purposes - especially for communication about our work. So it does matter that we have an up-to-date, reasonably standardized alphabet for our papers and journals. But I also teach my students not to trust any alphabetic description of speech, and not to imagine that their or anyone else's transcriptions provide reliable descriptive units that capture all of the phenomena for which we linguists seek understanding.