I played a bit more with characterizing the voltage divider based on LT5400-5 resistor network to get an extra 12V range on HP3457.
The chip that I got gives exactly 4:1 ratio as well as I can measure, within about +/-2 ppm.
The noise of the meter on the 3V range with the divider connected is increased just by the Johnson noise of the divider.
10 PLC: short circuit: 280nV rms, divider shorted: 352 nV rms, difference: 210 nV rms
100 PLC: short circuit: 89 nV rms, divider shorted:109 nV rms, difference: 63 nV rms
The divider, when shorted, presents a 750 kOhm resistor, so the Johnson noise for 3 Hz BW (10 PLC) is 190 nV rms, for 0.3 Hz BW (100 PLC) is 60 nV rms. This matches very closely with the measured noise increase. So the current input noise of HP3457A is negligible.
Measuring the calibration of the 30V range on my meter relative to the 3 V range with a 2.5 V reference I found the 30 V range is higher by 2 ppm. After I correct for this, measuring a 10 V reference on the 30 V range and on the new 12 V range with the 4:1 divider connected to the 3 V range agrees within about 1 ppm, as shown in the plot.
The noise on the attached plot is not that good because I am using a cheap AD584 reference, which dominates the noise.
Next project in this metrology on smallest possible budget exercise: setup a battery voltage source.