[worktrunk-bot]

wt remove --force now saves a worktree's uncommitted changes before discarding them, using the same refs/wt-backup/<branch> ref and reflog that the squash backup in wt merge / wt step squash already uses. The output then prints the backup and a recovery command:

▲ Discarding uncommitted changes (--force):
  ?? devbox.lock
↳ Backup created @ refs/wt-backup/feature-untracked (690353c)
↳ To recover in a new worktree, run wt switch --create --base 690353c… --no-hooks recovery/690353c

Closes #4544.

What gets saved

• Staged, unstaged, and untracked files go into the snapshot, which is what git add -A would stage. The snapshot is built through a TempIndex, so the real index and the stash list are left alone.
• Ignored files, such as build output, are not saved. For submodules, only the checked-out commit is recorded; uncommitted changes inside a submodule are not saved.
• The backup commit's parent is the worktree's HEAD, so the branch tip also stays reachable after -D deletes the branch. On an unborn branch the commit has no parent.
• The backup is taken only when --force actually discards something. A clean --force removal creates no ref.

Decisions for review

• Always on, no flag. This follows the issue thread: a flag would grow the CLI, and always saving means -f loses nothing you can't get back.
• A failed backup refuses the removal. The error is Failed to back up uncommitted changes; worktree not removed. That follows the data-safety guidance in AGENTS.md, but it also means wt remove -f can no longer remove a worktree that git can't snapshot (for example, a corrupt index). The alternative is to warn and continue.
• Cost. Untracked files that aren't ignored are written as blobs. A large unignored artifact makes the removal slower and grows the object store until the reflog entry expires. Nothing prunes refs/wt-backup/ today, and this PR doesn't change that.

Code

create_safety_backup and the new create_removal_backup (src/git/repository/working_tree.rs) share a record_backup helper for the commit-tree / update-ref step. The helper looks up the branch name without the cache, and falls back to symbolic-ref so an unborn branch is still named. The removal side lives in stage_worktree_removal (src/git/remove.rs). That function runs right before the worktree is renamed into trash, so it covers wt remove, wt merge --remove, and the foreground path.

Tests

• test_remove_force_with_untracked_files now checks that the backup ref holds the untracked file's content, that its parent is the removed branch tip, and that an ignored file is excluded.
• A new unit test covers the unborn-branch case.
• The four remove_force_with_* snapshots and the wt remove --help snapshot gain the new lines. The docs (wt remove help, FAQ) mention the backup.

[worktrunk-bot]

Two small follow-ups, which I'll push as one commit: