Modelling SRPP driver with cathode feedback 7591 (7868 / 6GM5 / EL506) the grid resistor value of the output tube, the cathode resistors of the SRPP showed influence on the distortion cancellation. In the end the ECC81 showed to have better end results (speaker output) than the ECC83 yet looking at the SRPP on its own the ECC83 was a lot better. And then there is the differences between different versions of the driver tube long plate versus short plate, one brand vs another brand. The design was originally for SRPP with local feedback 7591/7868/EL506 (and no global feedback) but in the end a combination of local feedback and GNFB was decided upon to counter tube differences.
At one stage I had a PP amplifier with a floating paraphase inverter that I had made adjustable, i.e. one side could generate a larger (or smaller) amplitude than the other side. When adjusting for exactly the same voltage levels of the phase inverter the distortion (at the speakers) was higher than when adjusting for minimum distortion at the speakers. This was due to amplification differences between the output tubes (although they were matched).
At one stage I had a PP amplifier with a floating paraphase inverter that I had made adjustable, i.e. one side could generate a larger (or smaller) amplitude than the other side. When adjusting for exactly the same voltage levels of the phase inverter the distortion (at the speakers) was higher than when adjusting for minimum distortion at the speakers. This was due to amplification differences between the output tubes (although they were matched).
Real Matching powertubes is at least 2 parameters : current at some working point ( close to the actual
working conditions) and Gm ( amplification ) at the same working point.
What you described seems like a mismatched Gm
At any real load reflected by the OPT, one half of the entire audio power is lost/dissipated in the 3K resister's loading the 300Bs.
And large DC losses as well. Did you try your amp without the 3K resister's & driving straight into the OPT?
The parafeed in this case appears to offer no benefit at all.
870V in the chassis is dangerous to life.
In this case the mismatch was in sides of the 6SN7 driver that sat between phase inverter and the output tubes. What is matched anyway? In the past 10% was "close enough" and if you have a mismatch of 10% between the triodes of the 6SN7 and then 10% of the output tubes plus a slight difference between the windings in the output transformer then the floating paraphase can reduce distortion due to amplification by some adjustment.