fix(coord): count a session in a deregistered checkout as unplaceable

[wshallwshall]

The occupancy fence could not refuse on the one ground that matters. A session record whose cwd lies
inside this repo family but matches no registered worktree was dropped with a bare continue and
never reached $faults, so RecordsUnplaceable could not rise and Available could not go false
because of it.

That is the manufactured-evidence shape. Available: true was unfalsifiable on that ground, so it
proved nothing, and two proposed reapers are gated on it.

The control fired first

Against the unpatched script at b8421ca53, with only the new test added:

2 failed, 2 passed in 7.85s
exit: 1

AssertionError: a record naming a checkout of this repo that git no longer lists reached no counter
{'RepoFound': True, 'Available': True, 'RecordsExamined': 1, 'RecordsUnplaceable': 0, 'Sessions': []}
assert 0 >= 1

That receipt is the defect stated in its own output: the record existed, parsed, and reached no
counter. Patched:

4 passed in 5.08s
exit: 0

Neighbours, same venv, unpiped, exit echoed separately: 98 passed in 151.06s across
test_coord_presence, test_coord_claim_liveness, test_coord_box_activity, test_coord_recoverable,
test_coord_overlap_attribution, test_coord_handoff_report, test_session_context_presence and
test_worktree_remove. test_worktree_prune_merged.py: 79 passed in 247.29s. ruff format --check,
ruff check and mypy clean. HEAD was b8421ca53 before and after every run, so no run measured a
moving tree.

The boundary, and why it is drawn there

A cwd OUTSIDE the repo family is not a fault -- it is another repository's session. The fault is a cwd
INSIDE the family that matches no registered worktree.

cwd state	Verdict	Evidence
Inside a registered worktree	placed	never reaches the new code
Exists, its .git pointer resolves to THIS repo's common dir	FAULT	positive proof, and it survives deregistration
Exists, resolves elsewhere or nowhere	not a fault	another repo, or a plain folder
Gone, named <primary>-<slug>	FAULT	nothing on disk can say whose it was; the naming new.ps1 builds is the only evidence left
Gone, not named that way	not a fault

The name test is admitted ONLY when the directory is absent. A name match on a directory that exists
is the sibling-prefix trap that test_coord_presence.py::test_sibling_prefix_directory_is_not_treated_as_inside_the_repo
already pins, and MessageFoundry-vault sits beside the primary on this machine. Faulting it would
leave the fence permanently unavailable, which disarms every caller as thoroughly as never refusing.

Get-OwningGitCommonDir walks the filesystem rather than asking git, deliberately: git -C <path> rev-parse fails outright once a worktree's admin entry is pruned, which is precisely the state being
asked about.

Callers

All four already fail closed, so this can only move them toward refusal.

Caller	Change
prune-merged.ps1	one more route to EXIT_REFUSED; every candidate becomes SKIP. The fence is re-read once before the apply loop, not per candidate.
presence.ps1	stdout stays []; the UNAVAILABLE receipt goes to stderr. JSON contract untouched.
claim.ps1	Get-OccupancyOnce returns false, so a holder stays present, which already refuses.
remove.ps1, box-activity.ps1	read occupant lists, not Available. Unaffected.

No field or signature was removed.

Live reading, patched

6 config roots, 7 records, 0 unplaceable, 110 registered worktrees, 3 sessions placed, Available: True.
The new fault does not disable today's fence. That zero is reportable because the planted controls in
the test fire.

Two corrections to the brief behind this commit

The brief asked for a plant pointing at a deleted .claude/worktrees path. Those sit under a
REGISTERED worktree, so such a record always matches an ancestor and is placed. That plant would have
gone green against the unpatched script and proved nothing. The path has to be under a checkout git no
longer lists.

The brief also said Available could never go false. Too strong: three paths already set it false --
no config root, an unreadable registry, and zero readable records. All three mean "I could not look at
all". What could not happen was Available going false because a session could not be PLACED, which is
the only one that refuses on positive evidence.

Known follow-up, deliberately not in this branch

scripts/worktree/prune-merged.ps1's FENCE UNAVAILABLE header paragraph is an incomplete enumeration.
It omitted the cwd-less shape before this change and omits two of three after it. A separate session is
holding a patch for it, derived from Get-WorktreeOccupancy rather than from any prior comment, and is
waiting on this branch to land.