Thank you all for your replies.
I’ve done some more measurements, and I’ve also implemented an offset measurements based on RX timestamps while precisely setting the period between two TX packets via delayed transmission (as described here: Method #2 for tuning? - #3 by mciholas). This should serve as a somewhat independent measurement (and I don’t have a frequency counter or a spectrum analyzer available at the moment)
Here the raw RXTOFS and DRX_CAR_INT values as read from the chip and after performing sign-extension, each step is a +1 or -1 change in XTALT:
Here the calculated relative offsets between TX and RX (signs may not be correct as I just wanted to overlay the data to compare):
Measurements are for Channel 3, but they look similar on other channels. The TX was transmittig packets at a fixed period, while the XTALT on the RX was changed at +/-1 increment/step. The results look the same if I keep the RX tuning fix and change it on the TX.
All measurements have been taken at room temperature (fairly stable) and after letting the modules reaching something close to thermal equilibrium.
All three values agree quite well, so it looks to me the tuning resolution is in fact in the order of 3-3.5ppm/step, no matter what the datasheet states.