Yet in ways, this paper seems like a confession that he's reached something of a brick wall with the problem.
If you look at the bidirectional link, http://lambda-the-ultimate.org/node/1526, he goes over a series of partial solutions before discussing his (very general) approach. I suspect that a workable system might be better through putting together all the classes of partial-but-well-behaved solutions rather than attempting a fully general solution.
Another thing that interests me is bidirectional code that operates on itself.
If "the problem" is that of providing accurate, fully static types for this type of transform, then of course yes. ;)
I think the end result is improved (easier to learn and more flexible) for consciously relaxing that goal, however.
(Edit to muse: I think 'relaxing' is a far more intelligent & interesting approach to overly-ambitious goals, as opposed to 'abandoning'.)
In particular something like this will (I think) be a perfect substrate for refactoring parsers, semantic diff tools, and other neat stuff...