Previously, the protocol used a mix of strings and bytes without any consistency.
When we go to multiple platforms, we won't be able to mix encodings of paths anyway.
We don't know this is the right approach, but it at least makes things consistent
and easy to read in the database, on the wire, etc. Really, we should be using entry
ids etc to refer to entries on the wire anyway, but there's a chance this is the
wrong decision.
Co-Authored-By: Nathan Sobo <nathan@zed.dev>
The new version of rust-analyzer changed the disk-based diagnostic token
to `rust-analyzer/checkOnSave/0`. The trailing number could be different
from 0 when there are multiple Rust projects open using the same rust-analyzer
instance.
As such, with this commit we will perform a prefix match as opposed to a strict
equality check when detecting a disk-based diagnostics progress token.
This fixes an issue with the Go language server, which reports invalid
formatting ranges when there's a missing newline at the end of the file.
Specifically, if the buffer is `N` lines long, it will try to insert the
newline at `Point(N + 1, 0)`.
I confirmed the behavior is the same in VS Code, and they indeed clip the
LSP ranges as well.
