Is anything blocking this patch? @alexbiehl @harpocrates ?

Sorry, just me not noticing the notifications..

This fix doesn't seem to work on Windows. I'm afraid that even with verbose output, there are no meaningful error messages (I've attached the output anyway).
(Note that I had to patch GHC 8.6.1 for plugins to work on Windows, see GHC trac #15700.)

cabal new-haddock output, with GHC 8.6.1

This fix doesn't seem to work on Windows. I'm afraid that even with verbose output, there are no meaningful error messages (I've attached the output anyway).

Yeah, the -v3 is something that cabal doesn't end up threading to Haddock (perhaps it should...). However, I doubt that the verbose output is all that useful - the failure is surely happening in typecheckModule 😄.

Unfortunately I don't have a Windows machine with which to debug...

At first, I thought it boiled down to the fact that, when running haddock, the call to initializePlugins (from ghc/compiler/main/DynamicLoading.hs) determines that no loading needs to take place.
When running e.g. cabal new-repl or cabal new-build, the call to initializePlugins would reload plugins, whereas with cabal new-haddock, it would follow the code path which did not reload anything.

However, after modifying the code to force cabal new-haddock to always reload plugins, the problem persists. Let me know what I should be looking for in typecheckModule and I'll try to keep digging.