Okay, lot's of comments! Here's what I know (and as an audio engineer, look, we're not scientists - some of these arguments reach scientific levels to the point where it's no longer about what's important, which is perception). All I can comment on is what I've known for a while, and what other engineers concur with (we leave the scientific arguments to "audiophiles"). My responses to the key points:
"24 bit to MP3 involves two conversions... 24 bit to 16 and uncompressed to compressed." - No, it depends on what bit resolution the encoder can handle. LAME 3.99 can encode an MP3 directly from a 24bit file (the LAME encoder is one of the most, if not the most popular encoder in the mastering community, likely what most of the MP3s out there have been encoded with).
"By the way, VBR is better sound quality than CBR with compressed audio." - Perhaps, the idea of keeping things CBR is (from other engineers who share this POV) to keep the encoder consistent; VBR will reduce the size of the file even more and when creating MP3s, my idea is to keep as much of the data as possible (a bit counterintuitive, I suppose). Pondering whether VBR sounds better than CBR is right up there with saying one fake Sugar brand is better than the next; they're still NOT real sugar
"Secondly, several listening tests over at HA have shown that headphone reveal compression artifacts more easily than speakers" - This comment is a bit argumentative, I will avoid that aspect of this conversation, all I have for this comment is this: Not all headphone drivers are the same, and the same thing goes with Speakers. I don't know what was used for these listening tests, the room (who designed it) and how neutral the environment was, all I know is the equipment I use, whose purpose is there for wide-ranging translation of the material I work on to whatever my clients listen to, and besides, who is "HA"!? (LOL, don't answer)
"Anything to back up that claim?" - Just a little over 15 years of experience working with loads of recording/mixing/mastering issues from various sources (no science degree here, sorry). Trust me, the only ones obsessing over things like ABX testing aren't typically the people responsible for delivering this content to the masses (tends to be really analytical audiophiles that love to argue more than just chilling out and listening to the tunes!)
"MP3 320 CBR doesn't filter out stuff above 16 kHz. As someone who masters audio you should know that." - I mentioned the LAME encoder, other encoders might not do it, but I mentioned that LAME is one of *the* most popular encoders, so likely some commercial tracks that people download have everything past 16kHz filtered out. Versions prior to 3.99 didn't, but I see a lot of MP3 files with missing frequencies past 16kHz; it's not that uncommon.
"The MP3 psychoacoustic model doesn't remove frequencies which are more likely to be "felt" than "heard". - On paper, this is the truth. In the real world, I'll tell you what I did one time. I grabbed my wife's favorite CD and made a 320kbps MP3 version of it and I popped it in the car stereo on our way to run errands. After about 10 minutes, she said "Something's wrong with this CD, it sounds like it's going bad or something." My wife completely hates anything to do with analyzing audio and is simply someone who loves music without having any affection to anything technical (my favorite kind of person). Once you start talking science, it's possible to start believing things because someone can present a very articulate reason to the cause; I don't care for written theories of how a person perceives sound.
"I'm suspecting his problem is when he's bouncing from 24 to 16. The wrong dither can indeed make a huge difference. If he's exporting from a 24 bit ProTools mix directly to MP3, ProTools does the downconvert in the background using a default dither." - The Pro Tools dither not the best, that's for sure. Again, LAME allows direct encoding from 24 bit files and I personally batch process MP3 files (or simply have the clients send whatever the digital distribution house needs to create their metadata-encoded MP3s for iTunes, etc.) 24bit masters printed into PT are exported without any other processing from my end. I often don't create MP3s for clients, this is often done by distributors and once in a while, I load the MP3s into my system and can tell that there's a huge difference, it's almost like fat free vs. the real whatever it is you want to compare fake stuff to.