# Reproduce the twenty Git cases

Requirements: Git 2.56.0 and Python 3.12 or later. No Python dependencies are required. Use a Git installation with the expected version; the scripts refuse other versions. All work happens in fresh temporary repositories and is removed afterward. No repository in the current directory is opened or changed.

Run every case:

```sh
python3 matrix.py --git git --output rerun.json
```

Run one case:

```sh
python3 case-01.py --git git --output case-01.json
```

Run the additional release checks:

```sh
python3 release_checks.py --git git --output release-checks-rerun.json
```

If Git is not on PATH, pass its executable as the --git value. That value is used privately and is not written into the result. The configured identity is fictional, TZ is UTC, and both Git dates are fixed. Output paths in command logs are repository-relative; linked worktrees use a relative sibling path.

The twenty case files are entry points into the original matrix.py implementation. The published runs/ directory holds each current observation. R03, R09, R11, R16 and R20 were rerun for this revision; the other fifteen records retain their earlier observation times. Match before/after state, command output and assertions; ignore observed_at when comparing a new run. The record’s logical index hash deliberately excludes timestamp caches so state comparisons are reproducible.

R03 demonstrates replacement of a staged-only version by HEAD. R09 and R11 retain a staged/unstaged split in the stash but show that plain apply does not reinstate staging. R16 quotes the reflog expression. R20 uses ../topic-worktree without enabling relative-worktree configuration. Review each starting state and warning before adapting any command outside these disposable repositories.

The release checks include text, binary and modify/delete examples of the marker-based --resolved option. Binary and delete conflicts need deliberate resolution; the command can stage them unchanged. Explicit paths limit the selection, and git diff --cached is checked before any proposed commit. A separate clean-repository comparison shows apply --index restoring staging where plain apply does not. These are declared synthetic cases, not coverage of every conflict shape.

The source code is MIT licensed. Original observation records are CC BY 4.0. No Git executable or third-party package is included.