Might do this for everything but let's start here. I initially went
with backslashes for URLs because that's how Vim on win32 normalizes
buffer names, and figured it might simplify things like equality checks.
But Vim itself breaks in some places. Editing a `file://` URL doesn't
work if backslashes are in use, and Vim unescapes incorrectly when
constructing <q-args> (see 6356bbc4a7).
Work trees and submodules have two things that could be called the Git
dir: the directory itself, and the ".git" at the root of the work tree.
Introduce "fugitive_dir" as our name for the one we consider canonical.
This is still a garbage fire, but at least now the use of `.git/refs/`
paths has been relegated to the public interface. This also fixes a lot
of quirks, and hopefully doesn't introduce too many new ones.
This will enable functions like FugitiveFind() to use fugitive#repo() as
the optional second parameter, which should make transitioning to the
new API a bit easier to do incrementally.
The \s and \S atoms consider space and tabs to be the only valid
whitespace characters, while the [:space:] character class also includes
control characters like newline and form feed.
This is not and can never be 100% perfect. There's no way to work back
from a file like info/exclude COMMIT_EDITMSG to the work tree. So
core.worktree continues to be recommended.
References: https://github.com/tpope/vim-fugitive/issues/1920
* Don't attempt to handle relative paths, as there's no guarantee the
current working directory is the one Vim was started with. In
practice, the only relative path I've seen is `.git/index`, which is
already the default and thus harmless to ignore.
* Cache the result of FugitiveVimPath(), to allow for slow
implementations.
FugitiveVimPath() is intended to potentially convert between UNIX and
Windows paths in mixed environments. Let's separate uses that require
that from those that simply normalize slashes.
The 2 slash version slots the drive letter into the host field, which
I'm worried will cause problems when the URLs are used non-opaquely, for
example, with an LSP. Let's start transitioning to the convention used
by file:// URLs.
Using $GIT_DIR/index for the summary buffer has 2 problems:
* It requires a BufReadCmd for all files named "index", necessitating
special handling for false positives.
* It forces us to resolve ".git" files and symlinks, decoupling us from
the worktree and thus forcing us to depend on `core.worktree`. Git
always sets this when necessary, but users and third-party tooling
sometimes do not.
Using a fugitive:// URL for the buffer solves both.
This is a large, breaking change, so let's leave $GIT_DIR/index as the
default for now.
