I agree with you, Paul. Deconvolution works great for well-defined optical systems such as a microscope or telescope where one uses a fixed focal length lens with an extremely shallow depth-of-field and extensive corrections for optical aberrations. Deconvolution has been used for those systems for 25 years that I know and their widespread use probably dates back to the early development of computers when the capability to perform the necessary computations became practical.
With typical cameras that most photographers use, I think it would be difficult to use deconvolution without introducing lots of computational artifacts since the optical system is ill-defined. Think about how drastically the optics change when zoom lenses are used and lens focal length is a huge variable. When we add the variable of focus distance which can vary a lot for the large depth-of-field lenses which we use, we have another ill-defined variable to consider. Finally, our typical camera lenses are poorly corrected compared to microscope and telescope lenses so optical aberrations vary a lot within our image area.
Considering the optical unknowns for our typical photography systems, it may be possible at best to target one or two problems which could be improved by deconvolution but I think a lot of other problems are present which throw a monkey wrench into it. As you said, unintended negative side effects seem likely to appear in the computations. I suspect that deconvolution for typical camera systems that we use are unlikely to be routinely applicable to all images but may be useful in specific instances.