Without further ado here is the schematic. Description of what's going on is below.

The biggest change is the use of a CCS diode, CR1 to isolate the pentode's grid-cathode voltage variations from the voltage reference. The other changes are setting up the noise filter for the voltage reference as a active filter and adding R13.

I had found that the biggest  leakage path was through the voltage reference. In the cascode configuration the pentode does a good job of isolating the rest of the circuit from the AC voltage variations of the signal that is being amplified. However, the pentode itself generates an AC voltage because it's grid bias changes as the voltage across the pentode changes. This voltage change is coupled to the screen by C1. The voltage variation is then coupled into R4 and causes the current through the voltage reference circuit to vary. This leakage limited the effectiveness of the CCS to about 65db of isolation. By adding the CCS diode CR1 (1N5294) the voltage reference is now fed by a CCS. This increased the isolation to over 85db. 

Earlier experiments to contain the leakage current included a cap, a zener and a 431 shunt reg. For these experiments R4 was split into 2 resistors R4a and R4b, the 100K remained on top and a 18K was added between R4a and U1 (the location in the schematic where CR1 is).

In the first attempt a 5uf capacitor was connected between the junction of R4a and R4B and the source (cathode for you tube folks) of the mosfet. This produced the highest isolation I have measured so far- 92.5db. Unfortunately it did not sound good. The cap caused the low level details to be smeared. It also required another cap, something I didn't want.

The second attempt was a zener diode. One advantage was it worked all the way to DC. Sound was much better, the transparency and detail was better than the Rev 2 circuit. The down side is zener diodes at low current are really noisy. The noise added by the zeners was noticeable when you were within a foot of the speakers.

The third attempt used a 431 shunt reg setup like the one in the voltage reference circuit. This worked great. Had the same sound as the zener circuit but did not have the added noise.

The current version that beat out all the others is the CCS diode. It removed another layer of grunge that I didn't know was there. I suspect the reason is the mosfet is not having to deal with the leakage current. Even though the mosfet has very low output impedance (<200 ohms in the driver stage and <25 ohms in the output stage) and R4 was 100K ohms the interaction was still audible.

With the CCS diode feeding the voltage reference there is no need for the shunt circuits to control the leakage current. The impedance of the 1N5294 diode is speced at 1.15 meg ohms. With this being 10 times more than the 100K resistor I have been using the improvement should be about 20db.  Measurements confirmed that was indeed happening. The sound quality did improve. The back ground went back even further, the low level details are more obvious. The .75ma version was chosen as a compromise (isn't everything?). The 431s are speced at 1ma minimum. When you look at the performance curves they perform pretty good down to .5ma with one of the manufacturers actually specing them at .5ma. With the CCS diodes, the lower the current the higher the impedance.   I chose .75ma to get the highest impedance that I could thinking that  431s should still work well at that operating current. So far so good. The 91 volt zener is there to protect the CCS diode during startup and shutdown. The CCS diodes have a minimum breakdown voltage of 100 volts.

The idea for the active filter setup for the noise filter came after I was looking at Steve Benches active crossover schematic. I realized that with the mosfet in the circuit I had all the parts necessary for a 2 pole active filter. The advantages to this are better filtering and it uses much smaller caps. With the 240K resistors and .22uf for the first cap and .1uf for the second cap it gives a roll off point of 3.3Hz with a Q of 1. In this circuit configuration only the .1uf needs to be super audio quality. This gets into a more realistic range where the caps don't cost as much and Teflon caps are reasonable.

There is a new resistor added to the circuit between the mosfet and the pentode. The purpose of this is to keep the pentode out of A2 operation. When the CCS is run into hard clipping the mosfet keeps wanting it's current. It will keep working down to less than .5 volts across it. The turn on voltage of the mosfet is 3 to 4 volts. This means that the mosfet will can try to apply up to 2.5 volts positive grid bias to the pentode. This causes the pentode to draw grid current. The grid current drawn pulls down the voltage on C2 and reduces the current the CCS is sourcing. Recovery time is several seconds, not a good thing. By adding R13 and selecting the value so that the voltage it drops is .5 volts more than the gate to source (grid to cathode) of the mosfet, the mosfet will saturate first and keep the pentode out of A2 operation. Recovery now is instantaneous as the voltage reference isn't affected. R13 is needed only where the CCS will be driven into clipping. Output stages and/or driver stages that don't have much headroom.