Completions should not depend on Typecheck

[pepeiborra]

The ProduceCompletions rule depends on TypeCheck, which means it cannot be computed until typechecking is done.

While this can be addressed by using stale data, I think it would be more elegant to improve the implementation so that it only needs the parsed module + the ModIface of (the transitive closure of) imports and the package database.

Additionally, we could break down ProduceCompletions in two parts:

1. ProduceCompletionsFromImports
2. ProduceCompletionsFromLocalModule

The first one would have no dependency on the ParsedModule but only on the ModSummary, which is already smart enough to survive across edits that do not alter the module import list, and therefore only ProduceCompletionsFromLocalModule would need to be recomputed across edits.

[ndmitchell]

One day Haskell will have working records, and on that day, completions will require the type checked module. Your plan doesn't seem unreasonable for now.

[pepeiborra]

Currently completions are not type-directed at all, which is why I think depending on typechecking is unnecessary. When record-dot-syntax lands the completion engine will need to be enhanced to support it, and we can make it depend on typechecking info when a record field is being completed, and only then.