Rather than think of something original to say, I'm mostly going to respond to everybody else's ideas. It sounds like a lot of us agree on the basic idea, which is that performance-related data and audio-quality data might be better stored at different levels.
On the word "recording"
I agree with mudcrow. To me the word "recording" will always mean the original act of performance/recording, as in "what year did the Rutles record their first album?"
Of course, that's pretty firmly rooted in the idea of live or in-studio performances. Maybe it doesn't work as well for the more modern conventions of remixes, re-use and appropriation?
I'm glad we're looking at different terminology as part of the solution.
On the number of levels
I agree with reosarevok. Further levels should only be created if doing so won't create (even more) confusion.
caller#6's use case
Miles Davis recorded 'round Midnight at least 35 times.
MB currently has 136 "recordings" of 'round Midnight by Miles Davis. 
I might have the time to add all the performance ARs to the 35 session/recordings before I die. I definitely will never have time to add duplicates for each of the remaining 101 tracks, and there'll be more reissues and compilations every year.
This is not an edge case in jazz.
As a MusicBrainz user, I'm interested in finding the 35 session/performance/recordings. If I needed to know which releases have decent audio-quality, I'd check the reviews at amazon or at an audiophile site.
I suspect that's the BBC position as well, and the reason they're asking for recording-groups.
What does caller#6 want?
A "mix" as described by LordSputnik and nikki sounds good.
If we had a "master" entity, and if master:track was 1:n (excluding mashups or whatever), then won't we have the same questions? (i.e. "When should masters be merged?") How many editors have the equipment, the trained ear, the access to multiple files needed to make those edits?
If so, Kuno's tracks-with-MBIDs make more sense to me. No merging. Every "manifestation" is unique. (Although, to be honest, either masters or tracks-with-MBIDs is fine with me because I'll never be the one merging them anyway.)
I also like Kuno's framing of the question in terms of the FRBR model.
Having both masters and tracks-with-MBIDs (as LordSputnik suggests)? I'm not sure what advantage that has over a three-tiered system.