diff --git a/scripts/coord/overlap.ps1 b/scripts/coord/overlap.ps1
index b79cc9f9..7e9a4a81 100644
--- a/scripts/coord/overlap.ps1
+++ b/scripts/coord/overlap.ps1
@@ -260,7 +260,14 @@ if (-not $Refresh -and (Test-Path -LiteralPath $cacheFile)) {
     } catch { $map = $null }
 }
 if ($null -eq $map) {
-    $map = Build-Map
+    # @() for the SAME reason Write-JsonArray filters nulls, at the other end of the same round trip.
+    # A bare `$map = Build-Map` survives in process -- assignment preserves AutomationNull, so the
+    # zero-rows guard below still fires -- but it serializes into the cache as `"rows": null`, and the
+    # next run reads that back as `@($null)`, a ONE-element array holding $null. Count is then 1, the
+    # guard does not fire, and the table prints a phantom occupant. The hook arms this constantly:
+    # collision_gate.ps1 runs `overlap.ps1 -File ... -Json` on every gated edit and the cache is written
+    # before the -File early exit, so any bare `overlap.ps1` inside $CacheSeconds inherits it.
+    $map = @(Build-Map)
     try {
         New-Item -ItemType Directory -Force -Path (Split-Path $cacheFile) | Out-Null
         # Last-write-wins on purpose: a duplicate walk is the only cost of a race, and a lock on the
diff --git a/tests/test_coord_overlap_cache.py b/tests/test_coord_overlap_cache.py
new file mode 100644
index 00000000..7375b305
--- /dev/null
+++ b/tests/test_coord_overlap_cache.py
@@ -0,0 +1,189 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# Copyright (C) 2026 MessageFoundry Organization and contributors
+"""The overlap cache must round-trip ZERO rows as zero rows.
+
+``scripts/coord/overlap.ps1`` writes its walk to a per-repo cache and every later run inside
+``-CacheSeconds`` reads it back instead of re-walking. The empty answer does not survive that trip by
+itself. ``Build-Map`` ends in ``return $rows`` over ``$rows = @()``, and PowerShell UNROLLS an empty
+array on the way out, so a zero-row walk yields AutomationNull. In process that is harmless -- the
+zero-rows guard still fires -- but ``ConvertTo-Json`` serialises it as ``"rows": null``, and the next
+run reads that back with ``@($c.rows)``, which is a ONE-element array holding ``$null``. ``Count`` is
+then 1, the guard does not fire, and the human report prints a PHANTOM OCCUPANT: a blank worktree
+name on a blank branch with "0 changed file(s)".
+
+This is the same class of defect as the ``Write-JsonArray`` null filter above it in the same script,
+at the other end of the same round trip, and the same class as the ``Get-WiredMatchers`` unroll in
+``scripts/worktree/install-gate.ps1`` -- an enumerable returned from a PowerShell function is not the
+value the author wrote.
+
+WHY IT MATTERS RATHER THAN MERELY BEING UNTIDY. ``collision_gate.ps1`` runs ``overlap.ps1 -File ...
+-Json`` on every gated edit, and the cache is written BEFORE the ``-File`` early exit. So the hot path
+arms the cache constantly, and any bare ``overlap.ps1`` run inside the window inherits it: a session
+asking "who else is in this repo" is answered with an occupant that does not exist. An invented
+collision is not a safe direction to fail in -- it is the answer people work around, and a report that
+cries wolf on a quiet repo is one they stop reading, which is how a REAL collision goes unnoticed.
+
+Driven against real throwaway git repos, because the question is entirely about what the script writes
+to disk and reads back. A test over stub rows would assert only that a value someone else computed
+gets carried, and the value is exactly what is wrong here.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import subprocess
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+OVERLAP = ROOT / "scripts" / "coord" / "overlap.ps1"
+TIMEOUT = 60
+
+pytestmark = pytest.mark.skipif(
+    shutil.which("pwsh") is None or os.name != "nt",
+    reason="overlap.ps1 needs pwsh on Windows",
+)
+
+
+def _git(repo: Path, *args: str) -> str:
+    proc = subprocess.run(
+        ["git", "-C", str(repo), *args], capture_output=True, text=True, timeout=TIMEOUT, check=True
+    )
+    return proc.stdout
+
+
+def _repo(root: Path, name: str) -> Path:
+    """A repo with an ``origin/main`` to diff against, and nothing else in it."""
+    origin = root / f"{name}.git"
+    subprocess.run(
+        ["git", "init", "-q", "--bare", "-b", "main", str(origin)], check=True, capture_output=True
+    )
+    primary = root / name
+    primary.mkdir()
+    subprocess.run(
+        ["git", "init", "-q", "-b", "main", str(primary)], check=True, capture_output=True
+    )
+    _git(primary, "config", "user.email", "t@example.invalid")
+    _git(primary, "config", "user.name", "t")
+    (primary / "alpha.txt").write_text("base\n", encoding="utf-8")
+    _git(primary, "add", "-A")
+    _git(primary, "commit", "-qm", "base")
+    _git(primary, "remote", "add", "origin", str(origin))
+    _git(primary, "push", "-q", "origin", "main")
+    return primary
+
+
+def _cache_path(repo: Path) -> Path:
+    common = _git(repo, "rev-parse", "--path-format=absolute", "--git-common-dir").strip()
+    return Path(common) / "mefor-coord" / "overlap-cache.json"
+
+
+def _run(repo: Path, sandbox: Path, *args: str) -> subprocess.CompletedProcess[str]:
+    """Drive overlap.ps1 against one throwaway repo, with every machine-global input redirected.
+
+    ``-ConfigRoot`` and ``-TasksDir`` point at paths that do not exist on purpose: without them the
+    script reads this developer's REAL session registry and task lists, and the answer would depend on
+    who else is logged in. They cannot contribute a row here anyway (only worktrees of this repo are
+    walked), but a fixture whose result varies with the machine is not a fixture.
+    """
+    proc = subprocess.run(
+        [
+            "pwsh",
+            "-NoProfile",
+            "-NonInteractive",
+            "-File",
+            str(OVERLAP),
+            "-Repo",
+            str(repo),
+            "-ConfigRoot",
+            str(sandbox / "no-such-config"),
+            "-TasksDir",
+            str(sandbox / "no-such-tasks"),
+            *args,
+        ],
+        capture_output=True,
+        text=True,
+        timeout=TIMEOUT,
+        check=False,
+    )
+    assert proc.returncode == 0, f"overlap exited {proc.returncode}: {proc.stderr}\n{proc.stdout}"
+    return proc
+
+
+def _cached_rows(repo: Path) -> Any:
+    """The ``rows`` member exactly as it was serialised -- deliberately NOT normalised.
+
+    ``json.loads`` renders ``"rows": null`` as ``None`` and ``"rows": []`` as ``[]``, and the whole
+    defect is that PowerShell then treats those two as the same length. Coercing them here (``or []``,
+    ``len(rows or [])``) would reproduce the bug inside the instrument and the test would pass against
+    the defect it exists to pin.
+    """
+    return json.loads(_cache_path(repo).read_text(encoding="utf-8"))["rows"]
+
+
+def test_a_walk_that_finds_nobody_caches_an_empty_array_not_null(tmp_path: Path) -> None:
+    """Zero rows must reach the cache as ``[]``, and come back out as zero rows.
+
+    The positive control runs FIRST and against the same machinery: a repo that DOES have an occupant
+    must cache a non-empty ``rows``. Without it, an assertion that ``rows == []`` is satisfied just as
+    well by a script that never writes a row under any circumstances, or by a cache path this fixture
+    computed wrongly -- and a null result printed by a broken instrument is indistinguishable from a
+    null result printed by a working one.
+    """
+    # --- POSITIVE CONTROL: an occupied repo, so we know rows can be non-empty at all ---------------
+    occupied = _repo(tmp_path, "occupied")
+    peer = tmp_path / "occupied-peer"
+    _git(occupied, "worktree", "add", "-q", "-b", "peer-branch", str(peer))
+    (peer / "alpha.txt").write_text("peer edit\n", encoding="utf-8")
+
+    _run(occupied, tmp_path, "-Refresh", "-Json")
+    occupied_rows = _cached_rows(occupied)
+    # Deliberately shape-AGNOSTIC: `ConvertTo-Json` renders a single row as a bare object rather than
+    # a one-element array, and that shape is harmless -- `@($c.rows)` wraps it back into one row on
+    # the way in. Asserting a list here would make the control the first thing to fail under the
+    # defect below, which would report "the fixture is broken" for a fixture that is working.
+    assert occupied_rows not in (None, []), (
+        "the control repo has a peer worktree with a dirty file, so the cache must carry a row. It "
+        f"does not, so this fixture cannot tell an empty cache from a broken one: {occupied_rows!r}"
+    )
+
+    # --- THE CASE: a repo with no peer worktree and nothing dirty ----------------------------------
+    quiet = _repo(tmp_path, "quiet")
+    first = _run(quiet, tmp_path, "-Refresh", "-Json")
+    assert first.stdout.strip() == "[]", (
+        f"the walk itself must find nobody in this repo, or the rest of this test is moot: "
+        f"{first.stdout!r}"
+    )
+
+    rows = _cached_rows(quiet)
+    assert rows == [], (
+        "a walk that found nobody was serialised into the cache as something other than an empty "
+        "array. `null` here is the unroll: `return $rows` over an empty array yields AutomationNull, "
+        "and the next run reads it back as a one-element array holding $null, so the zero-rows guard "
+        f"does not fire and a phantom occupant is printed. Got: {rows!r}"
+    )
+
+    # --- and the read-back half: the cached answer must still be "nobody" --------------------------
+    before = _cache_path(quiet).read_bytes()
+    second = _run(quiet, tmp_path, "-CacheSeconds", "3600")
+    after = _cache_path(quiet).read_bytes()
+
+    # Guard the guard: prove this run was a cache HIT. A miss re-walks and REWRITES the cache with a
+    # fresh `at` stamp, so unchanged bytes are the evidence that the cached value is what was read --
+    # otherwise a passing assertion below would only mean the second walk also found nobody, which
+    # says nothing about the round trip.
+    assert after == before, (
+        "the second run rewrote the cache, so it re-walked instead of reading what the first run "
+        "stored -- this assertion would then be testing the walk twice and the cache never."
+    )
+    assert "No other worktree has changes." in second.stdout, (
+        "reading the cached empty walk back produced an occupant. Under the unroll the cache holds "
+        f'`"rows": null`, which `@($c.rows)` turns into one $null row.\n{second.stdout!r}'
+    )
+    assert "changed file(s)" not in second.stdout, (
+        f"a phantom row was printed from the cache:\n{second.stdout!r}"
+    )
diff --git a/tests/tooling_manifest.txt b/tests/tooling_manifest.txt
index 9105f5cc..f0a85789 100644
--- a/tests/tooling_manifest.txt
+++ b/tests/tooling_manifest.txt
@@ -50,6 +50,7 @@ tests/test_coord_claim_refresh.py
 tests/test_coord_claim_release_history.py
 tests/test_coord_lock.py
 tests/test_coord_overlap_attribution.py
+tests/test_coord_overlap_cache.py
 tests/test_coord_overlap_signals.py
 tests/test_coord_presence.py
 tests/test_coord_seat_prompt.py