The feedback compensation cap is there as a placeholder until I see how the circuit really performs. In all the p-p circuits I've done, the transformer acts as dominant pole, actually rolling off signs of input/driver stage misbehavior so that I had to probe previous stages as well as the output to see what was really going on. There is no place for partial feedback here, unless one wants to do it right from plate to grid (or plate to follower) on the outputs. Having said that, I was just looking at on old copy of the ARRL handbook yesterday, where they show a capacitive divider used for "neutralization" (actually negative feedback) used on the final stage to quell oscillation. Food for thought, but only after I get the amp running in its current state.
The follower mosfets are selected for low gate charge and low input capacitance. The stage driving them is a triode with relatively low Rp. I don't anticipate any problems, and indeed have not had any with this arrangement in the past. The zeners are all the same type in the schematic (18V), except for the one regulating the negative bias(33V). They will not conduct unless the gate is overdriven. As for mosfets vs. BJTs, I don't like having to supply base current, or having to worry about recovery from saturation (perhaps a real concern in the case of the splitter). In the breadboard stage, I'll do a little tinkering with the input to provide maximum headroom for the splitter.
One interesting worry is how long it will take the cut-off follower to recover from a negative signal excursion. That's what square wave tests are for...
I went back last night and corrected the reference designators (and some values) to make everything line up nicely. I want to take a look at my notes at work to check on a few more values in the bias circuit, them I will post the revised schematic. The 6P14P data sheet claims that I need only around ~8V of negative bias for my planned quiescent current of ~30ma/tube. That's a pretty sensitive tube. We'll see what actually happens in the breadboard.