In parafeed, the modeling that I've done shows the plate choke inductance to be the most important. The object is to provide the tube with a load that does not fall too far below nominal, so that the tube distortion is not increased. This effect is independent of feedback, which reduces the intrinsic distortion but does not affect the causes of that distortion.
Here's my take on the reasoning. Parafeed with a really huge capacitor means the output transformer inductance in parallel with the plate choke's inductance dominates. That inductance will have a reactance that falls with frequency, and will present a low impedance to the tube at very low frequencies, leading to tube distortion. This is mostly a problem at high signal levels, where an ungapped parafeed output transformer will have large inductance, so the plate choke with its airgap will dominate.
Parafeed with a really small capacitor will present the tube with a capacitive reactance that increases as frequency decreases, so the tube is lightly loaded and has low distortion. The price you pay is reduced frequency extension. (This can actually be useful in the tweeter amp of a biamp setup, where the parafeed capacitor can be part of the crossover network.)
When you have the best balance, the tube sees a relatively constant load impedance, which is also largely resistive (not capacitive or inductive, i.e. a small phase angle) to the lowest possible frequency. Done right, the response and load impedance are better, to a lower frequency, than series feed with the same inductance. That's one of the technical advantages of parafeed.
This analysis is what has led me to recommend a capacitance of twice the plate choke inductance, divided by the load resistance - about 1uF for a 30H plate choke and 8K load. In the modeling, capacitors between half that value and twice that value are pretty good - the model is quite tolerant of modest variations in the parameters.
