Support Jujutsu (jj) colocated repositories

Assessment of running git-wipe inside a Jujutsu repository, verified empirically against jj 0.44.0 and git 2.55.0.

Colocated repos (jj's default since ~0.32) already mostly work: refs/heads are real git refs, git branch --merged, all six content strategies, fetch --prune, push --delete and branch -D all behave, and jj auto-imports a git-side branch deletion (Done importing changes from the underlying Git repo.), which is jj op undo-able. Four things are wrong or inert.

1. Detached HEAD leaks a phantom branch into local_branches()

src/git.rs:511-518 runs git branch --format=%(refname:short) and keeps every non-empty line. git branch always lists the detached-HEAD pseudo-entry, even with --format, so the result contains a bogus branch name.

• Plain git, detached HEAD → (HEAD detached at 45c62f9)
• jj colocated (HEAD is permanently detached at the working-copy parent) → (no branch), and it is locale-dependent: (aucune branche) under fr_FR

Consequences: the phantom name enters the merge-detection pool, so git cherry / patch-id probes are run against it and fail; it shows up in git wipe status; and git branch -D "(no branch)" would exit 1.

Note this is not jj-specific — it reproduces in plain git under any detached HEAD. merged_branches() is unaffected because parse_branch_list (src/git.rs:1106) already drops *-prefixed lines; only the --format path is missing that filter.

Fix direction: use git for-each-ref --format=%(refname:short) refs/heads (the same plumbing branches_with_gone_upstream and branch_committer_dates already use), which never emits a pseudo-entry and is locale-proof.

2. --delete-gone is permanently inert under jj

jj stores remote-bookmark tracking in .jj, never in git config — a colocated repo has no branch.<name>.remote / branch.<name>.merge entries (git config --get-regexp '^branch\.' exits 1). So branches_with_gone_upstream (src/git.rs:530) and branch_upstream (src/git.rs:455) always return nothing.

Not a bug so much as a limitation to document; jj users get no gone-branch detection. refs/remotes/* are written by jj, so merged-remote detection and remote deletion still work.

3. Current-branch protection is a no-op

current_branch (src/git.rs:406) returns the literal HEAD in a colocated jj repo. Nothing real is excluded from the deletion pool, so the bookmark at the working-copy parent is eligible. Low severity — such a branch is only deleted if proven merged, and jj keeps the commits reachable and the operation undoable — but the intended safety net is silently absent.

4. jj workspaces are not git worktrees

jj workspace add ../ws1 creates no git worktree; git worktree list --porcelain still reports only the main one. Worktree discovery, --min-age, lock handling and the worktrunk integration therefore have nothing to act on in a jj repo. Harmless, but the feature simply does not exist for jj users.

Plus: non-colocated repos

jj git init --no-colocate leaves no .git at all; git rev-parse --is-inside-work-tree fails and git-wipe exits via its is_inside_work_tree guard. Clean refusal, no data risk. Worth an explicit "colocated only" line in the README rather than any code change.

Start in src/git.rs at current_branch, local_branches, branches_with_gone_upstream, and branch_upstream, then review the existing for-each-ref callers. Verify detached-head behavior and jj colocated repositories against the documented cases, and update README for the colocated-only and unsupported jj features. Done means phantom branches are excluded without locale dependence and the limitations are explicit.