Drop a worktree's row when its directory is deleted without git

[DrewBradfordXYZ]

Two commits. The first is a plain bug fix. The second extends it to locked
worktrees, which is a policy choice; drop it if you'd rather follow git's
lock rule, and the first stands on its own.

The bug

Delete a linked worktree's directory directly (rm -rf, Finder, a cleanup
script) instead of with git worktree remove, and its row stays in the tree
for good, with no session in it. docs/how-it-works.md already promises that
"a worktree someone removed" disappears within a couple of seconds; for this
kind of removal it never did.

Commit 1: the fix

1. git worktree list still lists the checkout, marked prunable, until
   someone runs git worktree prune:

   worktree /path/repo-wt/feat
   HEAD 0c44b25…
   branch refs/heads/feat
   prunable gitdir file points to non-existent location
   

   parse_worktree_list ignored that line, so the reconcile saw the checkout
   as present (and nebula add would adopt it, through the same parse). It
   now leaves prunable stanzas out. The main checkout is never prunable.

2. The sync never ran. The 2 s loop only reconciles when
   worktree_probe_stamp changes, and deleting a checkout moves none of the
   mtimes it reads. The stamp now also carries the registry entries whose
   gitdir target no longer exists (git's own prunable test). That list is
   sorted and only changes when the filesystem does, so there's no re-sync
   every tick.

nebula still never runs git worktree prune, and a row that holds a session
or terminal is kept, as before. That row is #127's case: its session now
refuses to start and says why. This PR only covers the empty row #127 leaves
behind. One existing test,
remove_worktree_survives_manual_rm_rf, checked the prune through
list_worktrees, which now hides a prunable checkout either way; it reads
git's raw listing instead. The docs change is one clause in WORKTREE SYNC.

Commit 2: locked worktrees (your call)

Git never marks a locked worktree prunable, however gone it is; the lock
exists so a checkout on removable media keeps its registration. But Claude
Code locks the worktrees it creates and a killed session never unlocks (the
comment in remove_worktree notes this), so an agent's rm -rf'd worktree,
probably the most common case, keeps its row after commit 1.

Commit 2 makes list_worktrees also leave out a linked checkout whose .git
file is missing. Git's registry is untouched: nebula never unlocks or prunes.
The trade-off: an empty row for a checkout on an unmounted volume is now
dropped, and comes back as a new row (new id, default sort order) when the
volume returns. Before, it stayed.

How it was tested

• New tests, all passing; cargo test -p nebula-daemon is 335 passed on
  main (a00b760), 338 after commit 1 and 339 after commit 2.

  • Commit 1: parse_worktree_list_leaves_out_checkouts_deleted_without_git
    and reconcile_drops_a_row_whose_checkout_was_deleted_without_git fail
    before this PR. probe_changes_when_a_checkout_is_deleted_without_git reads
    the new stamp field, so it doesn't compile without it. (The fail-before runs were made against
    v0.42.0; the rebase onto main applied without conflicts.)
  • Commit 2: list_worktrees_leaves_out_a_locked_checkout_that_was_deleted,
    and the reconcile test also runs with the worktree locked. Both fail with
    commit 1 alone.
  • The reconcile test needs git 2.31+, when prunable appeared in porcelain
    output.

• Live, with cargo build -p nebula binaries (not an installed nebula), each
  run as an isolated daemon (NEBULA_RUNTIME_DIR / NEBULA_DATA_DIR /
  NEBULA_CONFIG_FILE), reading the worktrees table: add a project with one
  linked worktree, rm -rf it, wait 6 s, then touch .git/HEAD and wait 4 s.

  	before rm	6 s after rm	after a git touch
  v0.42.0 (before)	feat, main	feat, main	feat, main
  both commits	feat, main	main	main
  both commits, worktree locked	feat, main	main	main

  These live runs were made on the branch's v0.42.0 base; on main the same
  two commits apply unchanged.

• cargo fmt --check and cargo clippy --all-targets -- -D warnings are
  clean.

🤖 Generated with Claude Code