Well, a number of interesting comments have come in, but there is still room for more, especially as I can offer an example
It seems to me that you have a good handle on the process, though of course the discontinuities you show are frustrating. As it happens, I have a set of scanned images of a map that has rather less overlap in one direction than I would like. Even a second go did not eliminate all problems and there are areas in the final stitch where the original was bent at the edges and resolution and illumination fell away. I could partially, but never completely, eliminate them by masking and I am sure the best course would have been to add in extra scans that covered the difficult areas. However these admitted defects are not really relevant for this exercise and it sounds as though you already have your scanned images and have to make the best of them.
My interest was not so much in scanned images, though no doubt scanning is the best way to capture the detail, but rather maps photographed in sections with a hand-held camera mostly pointed as directly downward (i.e. normal to the map) as circumstances permitted. For this setup Hugin's translation model works well, with y, p and r to deal with the direction in which the camera points and X, Y and Z to deal with its position for each image. But scanning should simplify the situation by eliminating the need for optimising y and p - and so it proves. I attach a screen shot of the optimiser tab for my test case and a reduced image of the outcome, viewable here:
This optimisation is the most basic I could find and in practice I would probably add various refinements to improve the optimisation figure. Nonetheless, the average control-point difference in pixels is still only 0.9 and the maximum 2.0. That is at the optimum resolution determined by Hugin, which corresponds to an image of roughly 19000 × 12000.
Something needs to be done to anchor the angular orientation of the entire image. Here I simply set the roll of image 0 at 90 deg, to compensate for the orientation produced by the scanner I used. That value seems to work well enough. I see you rotated the image by hand. That also works, with care. However, normally I add horizontal and vertical control points, especially when, as in my example, there is a clear framing border. Allowing the angle of roll to be optimised with respect to some reference is very important for scanned images, where, as I think someone pointed out recently, some variation in the angle in which the original is positioned on the scanner seems inevitable.
There's no magic in the value I gave for the hfov, namely 20 deg. It's pretty arbitrary. But I think it is generally important not to allow it to be varied if Z is to be optimised. In the case of the normal camera image, you can vary either the fov or Z to compensate for zoom, but allowing both to vary can give problems.. I'm not sure what reality Z might represent for scanned images, but optimising it does seem to help significantly.
I am wary of using individual lenses if there is no clear difference in the lens, or its equivalent (i.e. the scanner). As my example shows, it is not necessary. My suspicion is that, though it sometimes does seem to be recommended for flat stitches, it is a hang-over from the time before translation parameters were included. I certainly used it myself. But I don't think, overall, it represents a set of lens distortion parameters that apply in the case of each individual image. Rather, it has the effect, at least partially, of simulating an angled lens, which in any case does not apply to scanned images. But I may be missing something. I wait to hear.
I wonder if using different lenses is at least part of the cause of your discontinuities. I do note that the two sections of plot boundary on either side of each discontinuity do not share a common direction and, indeed, as one moves from one discontinuity to the next, the relative angle between the sections seems to change. If you are to use individual lenses the normal recommendation is to have lots of control points widely spread across the whole image produced by that lens and that might be difficult where, as in your case, there is little overlap.
Of course, long linear features can offer control-point detectors difficulties if there are no distinguishing details on or near them. But in the case of your image showing the discontinuities the seam clearly passes near or even through plot numbers and they should give control-point detectors something to bite on. Thus you could try dragging a rectangle around them and selecting that great standby "create control points here": see e.g.
Or, of course, you could add them by hand, assuming they are not there already.
There is another possibility for mending these discontinuities, although I have to admit I have only used it once. I depends on the fact that the plot boundaries in your case look pretty straight near the discontinuities. Go to the control points tab and, for each of the two images between which a discontinuity shows, define a new line which uses the same number in the two images and, as far as possible, spans the discontinuity. Indeed, several of these spread along the band of discontinuities may be enough to push the optimiser into aligning the various pairs of sections.
Roger Broadie