The only base quality that matters is the source audio, pitch shifting will leave artifacts regardless, they could easily record each gear to sample the entire range of each which would be ideal, the problem lies in the way they synthesize the rev range, it sounds like the loop length is entirely too short which results in a huge loss of the overall tone and character of the source audio, you can try it by playing any audio through a sampler and changing the loop length so it's shorter and shorter, it will sound a lot like PD's engines, very lifeless and it loses almost all characteristics of the original audio. you can shorten the length to the point where it just becomes a drone or buzzing sound, not hard to figure out where they're going wrong.
Yeah, I said that.
Pitch shifting itself shouldn't really leave
artefacts (although practically no-one uses
sinc-function "interpolation" for non-native sample-rate playback, but that's probably overkill given the good that power could be used for), it's the blending of the next sample, shifted down, with the current one, shifted up, that does, because of constructive and destructive interference between the samples changing the sound of both.
Besides, pitch shifting samples is just physically wrong, except for Doppler, of course. What some games are doing is a more granular method, a bit like those time stretching effects you get with slow-motion effects in video editors, the ones that just slow the sound down without changing its apparent pitch.
Those are hard to do for general sounds, but since engine sounds are periodic and we know the engine speed, 'cause we're controlling it directly, you can literally play back only a short snippet of the sound, e.g. corresponding to a crank rotation, then repeat it a few times depending on how much "slower" it needs to sound, before moving onto the next snippet. For sped-up sounds, you miss a snippet or two out each time. There's a lot of fiddly technical stuff to avoid harsh clicks etc., but it's workable and a few of the big games use this, DiRT 3 being one for certain, and possibly RRRE.
The problem with that is the engine still sounds wrong in places where it's not sampled, still has sample-blending interference artefacts and potentially complicates the Doppler slightly (I think you just inject that into the snippet playback rate for a given sample, perhaps). I'm yet to work out if you can sidestep the "sounds wrong in places where it's not sampled" issue with a bit of "clever" consideration to the underlying engine configuration for pre-recorded samples, but it's perfectly possible with wholly synthesised sounds (with the problem that the base synthesis isn't easy to get sounding right).
I've also, for a few years now, been messing with a sampling system like LFS has, where you use a single, very short sample that corresponds to a crank rotation, and looping that. You actually have so much more fine control over it that there is no real lack of detail once you start modeling speed fluctuations etc. If you try to flatten the synthesised sound into a sample to be used elsewhere, it sounds awful, though, because that fine control isn't considered in the game, it's just expected to be in the sample (a big part of the reason iRacing sounds so good right now is the minuscule engine speed fluctuations arising from the drivetrain modeling - physical modeling appears to be PD's preferred method, too). You still have interference issues with these short samples, though, but I potentially have a way around that, too...
Given PD's sound guys don't do this in their spare time, and "simulation" is their job title, I'm fairly certain this is the sort of thing they've been "researching" "every day"; they're working on it. Probably not for PS3, mind you.