Following Daves hint, I removed the over voltage protection diodes D7 and D8.
By means of an Impedance Synthesizer circuit, I measured different resistors from 5MOhm to 50MOhm.
The gain calibration of the 121GW is now -6% low, that means, that the nominal test current of about 20nA is decreased by that amount also, that is -1.2nA.
These 1.2nA were leaked by D7, or by the difference in leakage currents between D7 and D8.
At 6V reverse voltage, I have determined leakage currents of about 500..800pA for the one diode, but about 3nA for the other, so this difference causes this effect, at least.
Without the diodes, I re-calibrated the 50MOhm and 5MOhm ranges, and measured the change of reading over a 3°C temperature change.
It's now -0.04% only, giving a virtual T.C. of about -133ppm/°C, instead of +6400ppm/°C, as with the diodes assembled.
That clearly indicates that the diodes were the root cause for the bad temperature behavior and out-of-spec of the 50MOhm range.
The replacement of the 1N4007s by a TVS (SM6T22CA) probably is much better for ESD / overvoltage protection, as these TVS react much faster than ordinary power diodes.
A TVS can't be used in that application, though, because they have leakage currents of up to 200nA @ 25°C, which is 10 times the reference current of the 50M range.
In other words, a TVS in place of D8 would simply short both the 5MOhm and 50MOhm ranges.
So a solution with two JFETs would be the correct measure, although I doubt that they'd withstand Joes ultra-violent HV pulse testing