Test conditions:
Rs = 3.3Ω
Rg = 1Ω
Rfb = 316Ω (X2)
Gain = 56 dB
Cross-coupled output gain 6 dB
Total Gain = 62 dB

ZTX851 Preamp NJM2068DD Gain 62dB Noise Floor
Results:
Output noise measures -80.96 dBu. (20 kHz BW)
Subtract 0.7 dB for converter loading and we get -80.26 dBu.
Referred to the input the Ein is -142.26 dBu.
That's an Rnv of 11Ω.
4.3Ω of that 11Ω is Rs and Rg.
Add another 4Ω estimated for 2X transistor rbb and that leaves 2.7Ω Rnv "unaccounted for."
NF is approximately 5.23 dB.
Noise density works out to be a about 0.43 nV√Hz.
The output noise is about the same at this gain using an NE5532, NJM2068 or OPA1612.
There's still some magnetically induced hum which is reduced by physical nulling.
It will undoubtedly be possible to actually pickup a fraction of a dB if this were in an actual box, not an open ProtoBoard so please don't degrade this into "where's the missing dB or Protoboards are shit."
I'm in no mood for it - it's time to move on with a layout.
I moved the servo sense point from the output of the cross-coupled stage to the outputs of the previous stage.
The reasoning for that is someone could run the board without having to install both THAT1240s.
In fact using the bridge circuit that defines CM voltage and sampling one stage upstream allows the board to run without any 1240s.
This also allows the servo op amp to be removed for troubleshooting or transistor matching.
Better that the differential Deboo, which has a common mode gain of "1" not be "blind" to the actual CM voltage and have to have its CM reference defined by the THAT1240s.
The disadvantage is that the servo does not correct the millivolt-level offsets of the THAT1240s themselves.
A jumper - providing options for both - might be worthwhile.