OK, I think that's looking pretty good from a fairly quick look!
The initial residuals don't look all the well described by a low order poly, so I'd never expect it to be amazing, so I think that's a pretty good start.
It would be fun (in another ticket) to build a LUT from the data, and then check that, when running on the same data, the residuals are basically zero, but I'm certainly pretty convinced by that last plot that the 3rd and 4th order polys are improving things by quite a bit.
Again, in another ticket, perhaps plot a) the RMS residual as a function of increasing poly order (including the pre-corrected value too), and b) the quality of fit as that increases, as it will start to be weird and unpredictable at high orders, I think.
Probably worth Robert Lupton casting his eye over this too, but I think this is a good start.