Initially, we (my collegue -- the "real" EE -- and me, the "crazy" physicist...) also questioned our setup and what errors we might have done when we found this discrepancy. Truth is, the BNC cables we used are good enough at the frequency range of that device that we shouldn't see any deviations of the delay. I checked the cables with my HP 8753C and they perform flawlessly up to > 1GHz delay-wise, above that some tiny wiggles can be observed, but still no significant deviation from the general phase delay, up to the 8753C's limit of 3GHz. Btw, the 8753 doesn't show the "funny" incline of the delay plot in logarithmic frequency mode.
Of course, the delay will be introduced by the length difference of the two cables, that was the intended effect. I also agree that the 4630B (like probably most VNAs) primarily evaluate phase (by measuring the balance between the I and Q signals) but can easily calculate phase delay by dividing continuous phase by (360°*frequency). If the DUT (cable's dielectric) hasn't got any significant dispersion in the observed frequency range and there aren't any resonances otherwise, this figure should be fairly constant as it actually is, provided the linear frequency scale is selected (see attached photo).
For comparison, I attached the phase measurement of our two cables as well, both with linear and with logarithmic frequency scale. These figures are both correct. To me, it actually seems that Anritsu erroneously used the frequency reading from the linear scale to calculate the delay in the logarithmic display, which would also explain that right at the end of the logarithmic delay plot, the reading drops back to the approximately correct range.
Nico, what's the firmware date and revision of your 4630B if i may ask?
Thanks and all the best,
Thomas
Edit: Added the delay plot with logarithmic time scale for better comparison.