This time, when completing the keyword (e.g. `fn` + whitespace).
The bug was actually a double-bug:
First, we did not resolve the impl in the macro-expanded file but in the real file, which of course cannot work.
Second, in analysis the whitespace was correlated with the `impl` and not the incomplete `fn`, which caused fake (where we insert an identifier after the whitespace) and real expansions to go out of sync, which failed analysis. The fix is to skip whitespaces in analysis.
It should be left biased, not right biased, because when e.g. the use has typed `h` then requested completion, the `h` is what we want to find, not the next token (which might indeed be inside a macro call).
I'm not sure why I wrote `right_biased()` to begin with (I remember I had a reason and not just "both should work"), I might've copied the code in `expand_and_analyze()` (which is wrong, because there it lookups on the speculative file, where right biased will always find the correct token and left biased not).
This is still not perfect, because there might not be an identifier already typed then we might still end up in a macro call, but this is the best we can do.
I.e. with `fn foo()`, don't complete at `x.fo|`, but complete (with imports) for `x.foo|`, since this is less likely to have false positives.
I opted to only do that for flyimport, even though for basic imports there can also be snippet completion (completing the params list for a method), since this is less universally applicable and seems not so useful.
To be accurate, only their methods are excluded, the trait themselves are still available.
I also excluded a bunch of std traits by default. Some less opinionated, like `AsRef`, which should never be used directly except in generic scenarios (and won't be excluded there), some more opinionated, like the ops traits, which I know some users sometimes want to use directly. Either way it's configurable.
It should be pretty easy to extend support to excluding only specific methods, but I didn't do that currently.
Traits configured to be excluded are resolved in each completion request from scratch. If this proves too expensive, it is easy enough to cache them in the DB.
Which caused the macros of the popular `tracing` crate to not offer completions.
The reason is rather complicated: it boils down to macro ignoring their input and completion always choosing the first expansion.
Prior to this commit we used to generate import paths, then zipped them with the existing qualifier to check if they agree on the path to import.
This is brittle when re-exports come into play causing items to have multiple applicable paths that refer to them.
This commit instead rewrites this logic by generating the import path for the qualifier, verifying that the rest of the qualifier resolves and then doing a final lookup on that resolution result for the final segment instead.
Previously all lints were assumed to be `#[warn]`, and we had a hand-coded list of `#[allow]` exceptions. Now the severity is autogenerated from rustdoc output.
Also support lints that change status between editions, and the `warnings` lint group.
This mainly aids in error recovery but also makes it a bit easier to handle lifetime resolution.
While doing so it also came apparent that we were not actually lowering lifetime outlives relationships within lifetime parameter declaration bounds, so this fixes that.
