HLS should always respond to resolve requests

[fendor]

We have multiple reports, where resolve requests (such as completion/resolve and codeAction/resolve) are rejected by HLS since the _data field of the respective LSP feature has not been populated by HLS. This makes sense, as we only support resolve for certain kinds of CodeAction/Completions, when they contain particularly expensive properties, such as documentation or non-local type signatures.

In #4463, I added a plaster for completions, deciding to always add the _data field, which allows our run-time system to resolve such requests to the completion plugin.
While this should work for completions, we have the same issue with completions provided by the hls-cabal-plugin, which serves completions for .cabal files.
Moreover, multiple plugins provide CodeActions but not necessarily CodeAction/resolve handlers, causing HLS to reject such requests.

This behaviour is troublesome, as vim clients (maybe emacs as well) seemingly show such rejections prominently in the UI, perhaps even outright not applying already present Codefixes.

The solution might be something like this:

If the _data field is Nothing we do not try to resolve a */resolve request to a specific plugin, but always return the request data unchanged. If the _data field is provided, we try to thread the */resolve request to the plugin which presumably handles the request.
We only reject the request, if the _data field is given, but no plugin owner can be determined.

This behaviour differs from the current behaviour only regarding how to handle _data = Nothing case, in which case we rejected the request.

Linked issues #4467, #4451

[fendor]

cc @michaelpj, @joyfulmantis, do you think this makes sense?

[michaelpj]

Yes, I agree with something like this. We didn't completely think this through: while we have differing levels of support for resolve depending on the plugin, the capability we advertise to the client is all-or-nothing. So clients quite reasonably expect to be able to resolve any completion/code action/etc. The spec doesn't say what you do if a resolve request fails, but I think "complain loudly" is not unreasonable behaviour for the client.

So what to do? I can see two options:

1. Be dumb and permissive: if no plugin wants to resolve a request, then just respond positively with the original item! Potentially this masks real issues, but I think it's not too bad: if a plugin thinks it can handle the request but it then fails to resolve it, we should still return a failure.
2. Try and be smart: something like what you suggest, where we try and figure out requests that we're "supposed" to resolve (e.g. those with a data field), and fail if no plugin wants to handle those. I think this is possible since we set data. So as long as we maintain the invariant that only things which need resolving get data, then it could be okay.

I think the implementation looks pretty similar: in both cases we'll need a catch-all handler, it's just that for option 2 it will decide whether or not to return an error depending on whether or not the item has a data field. So maybe it's easy enough to be a bit smarter and do 2?

[fendor]

I have a POC for (2), will send the PR tomorrow.

[Sivecano]

did the POC solve the issue?

[fendor]

@Sivecano Yes, now we just need a release with the changes, which likely will coincide with the release of GHC 9.6.7 and 9.12.2 (the latter was released last friday)