Dead reckoning can get quite accurate once you realise that cars drive on roads, so if you have a reasonably up-to-date map you can use turns and corners to "snap" back to the road and reset a good bunch of your accumulated error.
(There are two precision problems here — tyre diameter changes slighly while you're driving, but also it's not precise to begin with before you even turn on the car, due to tyre wear.)
You'd need to do something like calibrating wheel speed data while you have good GNSS reception, then you could use it for dead reckoning. But accelerometers and gyros are cheap…
P.S.: I didn't say wheel speed data isn't used, just that it wouldn't be precise enough on its own.
My understanding is that using plain accelerometers and gyros will drift quite soon due to noise, while some cars (e.g. Tesla) maintain their position quite well when doing to an underground parking structure and then coming back. So I actually do believe that they put a lot of weight to tracking wheels, which I believe would not accumulate that much error. (For more precise location they'd still need to use accelerometers to detect downhill/uphill, though.)
So I believe that the precision actually favors wheels, not IMUs.
In the underground parking case it doesn't actually matter too much if there's a certain % offset, because you're very likely driving a near-closed loop of some sort; and as you say, the wheel size ratio can be calibrated upon good GNSS data.
A friend tried to make a phone ruler app (using IMU) and the experience was that it didn't take very long for the distance to shoot to the horizon. So, thought I should try how this kind of solution would work with the hardware of today, but it turns out there seem to be zero such apps in the Play Store.. The tools of that category all seem to use a camera.
I think the lack of these tools is a testament to the approach not really working.