tree -> index diff for status

[Byron]

Based on #1350

---

diff-correctness → gix-status → gix reset

---

Improve gix status to the point where it's suitable for use in reset functionality.
Leads to a proper worktree reset implementation, eventually leading to a high-level reset similar to how git supports it.

Architecture

The reason this PR deals quite a bit with gix status is that for a safe implementation of reset() we need to be sure that the files we would want to touch don't don't carry modifications or are untracked files. In order to know what would need to be done, we have to diff the current-index with target-index. The set of files to touch can then be used to lookup information provided by git-status, like worktree modifications, index modifications, and untracked files, to know if we can proceed or not. Here is also where the reset-modes would affect the outcome, i.e. what to change and how.

This is a very modular approach which facilitates testing and understanding of what otherwise would be a very complex algorithm. Having a set of changes as output also allows to one day parallelize applying these changes.

This leaves us in a situation where the current checkout() implementation wants to become a fastpath for situations where the reset involves an empty tree as source (i.e. create everything and overwrite local changes).

Extra Tasks

Out-of-band tasks that just should finally be done, with potential for great impact.

• support for hasconfig as part of resolve_includes() without actual lookahead.

Tasks

• diff index with index to learn what we would want to do in the worktree, or alternatively,
  diff tree with index (with reverse-diff functionality to simulate diff of index with tree), for better performance as it
  would avoid having to allocate a whole index even though we are only interested in a diff.
  • Must include rename tracking.
• how to make diff results available from status with all transformations applied, to allow user to obtain diffs of any kind? Compare to tree-tree-diff which has that functionality already.
• update is_dirty() and Submodule::status() to do full status.

Status Enables

• cargo package and its use of complete status information.
• Add gitoxide backend Byron/built#1
• starship native dirty check (but needs diffstats of worktree-vs-index)
• built can get fully-functional is-dirty flag for 'describe()'

Inbetween

• sort out Oxidize Radicle-Git #1319

Next PR: Reset

• reset() that checks if it's allowed to perform a worktree modification is allowed, or if an entry should be skipped. That way we can postpone safety checks like --hard

Postponed

What follows is important for resets, but won't be needed for cargo worktree resets.

• a way to expand sparse dirs (but figure out if this is truly always necessary) - probably not, unless sparse dirs can be empty, but even then no expansion is needed
  • wire it up in gix index entries to optionally expand sparse entries
• gix status with implemented 'porcelain-v2` display mode

Research

• Ignored files are considered expandable and can be overwritten on reset
• How to integrate submodules - probably easy to answer once gix status can deal a little better with submodules. Even though in this case a lot of submodule-related information is needed for a complete reset, probably only doable by a higher-level caller which orchestrates it.
• How to deal with various modes like merge and keep? How to control refresh? Maybe partial (only the files we touch), and full, to also update the files we don't touch as part of status? Maybe it's part of status if that is run before.
• Worthwhile to make explicit the difference between git reset and git checkout in terms of HEAD modifications. With the former changing HEADs referent, and the latter changing HEAD itself.
• figure out how this relates to the current checkout() method as technically that's a reset --hard with optional overwrite check. Could it be rolled into one, with pathspec support added?
  • just keep them separate until it's clear that reset() performs just as well, which is unlikely as there is more overhead. But maybe it's not worth to maintain two versions over it. But if so, one should probably rename it.
• for git status: what about rename tracking? It's available for tree-diffs and quite complex on its own. Probably only needs HEAD-vs-index rename tracking. No, also can have worktree rename tracking, even though it's hard to imagine how this can be fast unless it's tightly integrated with untracked-files handling. This screams for a generalization of the tracking code though as the testing and implementation is complex, but should be generalisable.

Re-learn

• pathspecs normalize themselves to turn from any kind of specification into repo-root relative patterns.
• attribute/ignore file sources are naturally relative to the root of the repo, which remains relative (i.e. can be .. and that root will be always be used to open files like ../.gitignore, which is useful for display to the user)

[EliahKagan]

Regarding 91b6549, is it required to use PowerShell? Otherwise, shell: bash can be used, even on Windows, which gives Git Bash, so anything that is due to PowerShell rather than a POSIX-compatible shell being used may be fixable that way.

[Byron]

Thanks for the reminder! Yes, that would have worked too. Overall, I think using the ${{ env.var }} syntax is more portable and should be preferred. There might be other cases where it's definitely better to change the shell though.

[EliahKagan]

Yes, there is nothing wrong with ${{ env.var }}. The only reason I mentioned shell: bash was this bit from the commit message:

What's worse is that it doesn't seem to fail if something doesn't work as one
would expect.

Achieving the effect of set -e or set -u is less obvious and automatic in PowerShell than in POSIX-compatible shells, so I figured the ability to use bash might be relevant there.

[Byron]

Achieving the effect of set -e or set -u is less obvious and automatic in PowerShell than in POSIX-compatible shells, so I figured the ability to use bash might be relevant there.

And after writing the bit about powershell I also realized that I am so used to set -euo pipefail that maybe I forgot how bash would have done the same in that situation. But I am not sure how the defaults are set on CI, where one would hope a failing command indeed triggers failure by default.