Perhaps I should clarify as I wasn't really trying to argue as much as just add more thoughts into the discussion. As a plugin developer I generally think that having every plugin oversample internally is overall the better engineering trade-off (and sometimes you want different amounts of oversampling for different stages in a single plugin), but the real point I was trying to bring up is that when processing audio, the bandwidth beyond what's audible can often become meaningful due to distortion (including aliasing) products in non-linear processing. It's generally agreed that we only need about 20kHz of bandwidth as far as ears are concerned, but producing 20kHz of clean audio might require more bandwidth during processing.
The same applies to precision too. For distribution of end-user audio going past 16 bits is IMHO pure marketing snake-oil, but this doesn't mean that you might not want to use significantly higher precision when processing, because then numerical errors come into play. When considering feedback in IIR filters, even single-precision floats are sometimes insufficient and in extreme cases this might even lead to unstable algorithms. Sometimes you can fix this by choosing an alternative algorithm with better numerical behaviour, but in some cases you might opt to have your plugins work in double precision. Even simple volume control (and certainly sample rate conversion if you're playing 44.1kHz on a 48kHz DAC) does lose a tiny bit of numerical precision, so there's no harm in having the PC convert the 16 bit media to higher precision for mixdown before sending it to the DAC... but that doesn't mean there's any real benefit in distributing anything more than 16 bits.
Overall I feel like the biggest problem with a typical discussion (online or offline) with regards to audio quality, sampling rates and precision is that it's very common for people to read something that applies to processing and then insist that it also applies to distribution media where in fact the engineering tradeoffs for the two are very different. A fancy high gain guitar amp sim might want to process with 16x oversampling and use double-precision floats internally to improve the numerical behaviour of it's circuit simulation, but that absolutely doesn't mean that there's any benefit whatsoever in 700kHz/64bits for distribution. For whatever reason, people treat audio as if it was made of magic pixie dust (and that's really what I was talking about with the gold plated ethernet cables). In the real world there's no magic and digital audio really is just regular engineering and all about finding the right engineering trade-offs which tend to be different depending on the purpose.