Thanks for the replies, guys!
Paul, Zwift uses whatever the trainer is broadcasting, true. At the same time, Zwift is telling the trainer about the terrain. The trainer behaves very differently when going uphill and going downhill.
Zwift is trying to simulate downhill to some extent. If the hill is steep enough, you don’t have to pedal (at least on my trainer). So Zwift is in the loop.
My sample is quite short, unfortunately. About 5-6 minutes going uphill, and 1 minute going downhill. In my experience, a 40 watts drop will affect my heartrate well within a minute.
Russel, power in watts are given by the following formula:
Power = Torque x Cadence
An increase in cadence will result in an increase in watts, at the same torque, not a decrease.
What happened during my run was that I shifted to the lowest gear, which is not very low, mtb, and had to increase the cadence to keep the effort at the same level (and thus lowering the torque).
I think a good test would be to use watt-pedals. If the pedals show the same as Zwift, it’s all in my head. If not, there is something wrong somewhere.