Can't readily find info on PSRR tho'.
I suspect (well dodgy!) that as it's a 'modulated current source riding atop a second, matched current source' that PSRR will be reasonable...
We can break the PSRR into two parts - symmetric and anti-symmetric.
In the symmetric case, both rails drop the same amount, and by symmetry, for well matched devices, the output doesn't change.
In the anti-symmetric case, one drops and the other rises, and you are now exploring how flat the line of drain current at constant VGS is, versus the device transconductance. Because J112 is specified as a switching device, the datasheet is no help. Squinting at the 2SK170 datasheet, on the VGS = 0.3V line round VDS = 8V, makes me think 1V change of VDS will cause maybe 0.2mA of change in 2mA, and you would have something like 20mS of transconductance. So I think that would give you something like 10mV of output shift - so maybe 100:1 or 40dB.
This is very, very rough, but if we run with it, for arguments sake, if you want the total noise to be less than say 3uV, you need the rails to be cleaner than 300uV. The NE5534 solution is way overkill but would certainly do that. Something simpler might do.
A proper simulation would give a better answer, rather than eyeballing graphs for a completely different device!
If you are on linux, there are some good native options, look at gEda which I think comes with gspice.