I went back to #1 as I had quite forgotten the deployment scenario, just to see what additional "fun" you might need to be having - I thought this was going into a motorcycle (!) and was about to wonder about noise on the wires that would make trash out of all you careful testing and design.
Surprise. Not a motorcycle. Bigger surprise: Eighteen (18) pole switch. Times two. That wouldn't make working with the ESP32 a/d converter any easier. @madmark2150 suggested in #14 breaking it into several resistor chains, Im not sure how that would be done but sounds plausible.
Most of the below is still thinking there were 12 poles and but one switch.
as @DaveX has just said, shifting into the more linear area of the converter will improve things.
If the crappiness is consistent. A feature of this kind of inadequacy is that it might be better to not depend on it.
Which lands us back to the per-instance customization by some means as I have outlined in #56.
Or. Use 3 more pins and go digital:
This post on another thread describes a diode matrix decoder.
I would pull up instead of down, and not use a true 0 amongst the selectable switch points.
If you have the pins, can afford the diodes and don't mind all the extra wires (selling it hard here), we can show how to reconfigure the pictured circuit along those lines.
An 18 pole switch would need a 5 column diode matrix. So now 5 input pins and 7 wires to the switch.
Here may be a defensible case for multiple processors. Get a small cheap Arduino board for each rotary switch and make them into "smart" switches.
Hey, what's the fun of killing this if you miss out on opportunities to overkill it? 
a7