feat(worktree): report the trees prune-merged must never remove (BACKLOG #1294)

[wshallwshall]

Closes BACKLOG #1294.

prune-merged.ps1 now evaluates every registered worktree and emits REPORT-ONLY rows plus a
copy-pasteable command block for the ones outside its candidate set.

No -Apply path for them, no -Name override, and worktree_gate.ps1 is unchanged. The
exclusions stay exactly as they were -- .claude/worktrees is where EnterWorktree relocates a live
session, and this tool once removed an occupied worktree, deregistering it and then failing to delete
the directory, after which every git command in the working session failed. Only discovery is
automated; the human stays the actuator.

Measured motivation: 36 registered worktrees, 11 reachable by this tool. On 2026-08-19 the owner
removed 12 of the excluded ones by hand off a list assembled manually, because nothing produced it.

The risk model took three cuts, and the controls caught each

Cut 1 listed every idle clean tree under a "look finished" banner -- printing a removal command
beside rows reading 37 commit(s) not on origin/main. The content column contradicted the banner,
and the banner is what gets read.

Cut 2 withheld anything whose touched paths still differed from main. It failed its own
positive control: a branch whose content had landed, as a squash, was withheld because main's
copy of a shared file had moved on for unrelated reasons. In a repo where docs/BACKLOG.md changes
hourly that withholds everything, and an always-empty report is not a safe report -- it is an ignored
one.

Cut 3 asks the question that matters. git worktree remove does not delete branches:

• a tree on a branch keeps its commits through the ref after removal -- rung 2 of the
  recoverability ladder, so no content test is meaningful;
• a detached tree is the only one removal can strand -- rung 3, a race against gc -- so it is
  listed only when some other ref already contains its tip.

Fails closed throughout

An unavailable fence, an unreadable activity time, or an unreadable ref list all withhold. The
withheld count is printed and carried in the JSON so the omission is visible; the withheld
trees are not named, because naming them re-creates the suggestion the withholding exists to
avoid.

Occupancy uses Get-WorktreeOccupants from occupancy.ps1 rather than a hand-rolled filter. The
session field is WorktreePath; my first version guessed Worktree/Cwd, which matches nothing and
would have reported every tree as unoccupied -- the failure direction that suggests removals.

Tests

Three added, and the negative control was run on all three: neutering the collection reds every
one.

One of them originally passed with the feature completely broken, because it asserted a tree was
absent from a list that was empty. It now asserts the list is populated first, and its -IdleHours
was wrong for the fixture -- which is why that assertion fires.

78 passed. ruff check and ruff format --check clean on the changed test file.