There are actually two feedback loops, both sensing the second stage cathode voltage. The first feeds that voltage to the EF86 screen grid. This was in the original design, and is good enough for EF86s that meet their specified characteristics closely - i.e. mostly NOS from the better manufacturers. We use that in the Repro with a quiet LED in the cathode.
Our experience with the wider range of modern manufacture EF86s is that they vary much more widely, and many will not bias up properly in that circuit. So I added the transistor, whose base goes to a voltage divider on the cathode of the second stage. When that voltage reaches about 0.6v, the transistor conducts, and it controls the EF86 cathode current. That transistor has high impedance and must be bypassed; we use a low-voltage polymer electrolytic which has excellent high-frequency performance.
In both cases, the dominant pole of the negative feedback loop (which controls stability) is set by the second stage cathode bypass capacitor.
