Bikeshed-wise though I might recommend that these methods return the same error as most other methods, which is Result<u32, ErrorSelf::Error>. That would avoid the need for conversions and remembering to use the right .map_err at all use-sites.

Hm, in that case I would perhaps go in the other direction of getting rid of enum Error<U> entirely and would add some constructors to the Reencode trait to construct error variants of the user defined type for the few cases that Error has enums for today. The Error<Infallible> for example is particularly clunky. Sounds good?

I do agree Error<U> is pretty clunky, but one nice part is that you can stick in std::convert::Infallible and not have to worry much about it. With extra constructors that would require extra boilerplate on each impl I think?

Perhaps as a further refinement of your idea, a ReencodeError trait instead of methods directly on Reencode?

I updated this to use Error universally.

Ah that's a merge conflict with #2145 which just landed prior, I can help resolve too as well if you'd like.

If you don't find an opportunity to, I'll get to it tonight or tomorrow myself.

This is ready to go again now.