diff --git a/tests/challenge/test_store_notify_dir.py b/tests/challenge/test_store_notify_dir.py
new file mode 100644
index 0000000..3bc88f0
--- /dev/null
+++ b/tests/challenge/test_store_notify_dir.py
@@ -0,0 +1,114 @@
+"""Shared v2 notification directories: safe open/close, and the test session's leak cleanup."""
+
+from __future__ import annotations
+
+import shutil
+import socket
+import tempfile
+import threading
+import time
+from collections.abc import Iterator
+from pathlib import Path
+
+import pytest
+
+from hypertrain.challenge.store import ChallengeStore
+from hypertrain.data.store import LocalFSStore
+from notify_cleanup import remove_new_empty_notify_dirs
+
+from .conftest import PARAMS
+
+
+@pytest.fixture
+def root(monkeypatch: pytest.MonkeyPatch) -> Iterator[Path]:
+    # A private temp root, short enough for AF_UNIX paths (tmp_path is not), so cleanup is
+    # observed without touching other sessions' directories in the real one.
+    private = Path(tempfile.mkdtemp(prefix="htnd-"))
+    monkeypatch.setattr(tempfile, "tempdir", str(private))
+    yield private
+    shutil.rmtree(private)
+
+
+def _store(state: Path, objects: Path) -> ChallengeStore:
+    return ChallengeStore(state, PARAMS, None, None, lambda _: None, LocalFSStore(objects))
+
+
+def _started(state: Path, objects: Path) -> ChallengeStore:
+    store = _store(state, objects)
+    store._start_notifications_v2()
+    return store
+
+
+def _directory(store: ChallengeStore) -> Path:
+    assert store._ipc_path_v2 is not None
+    return store._ipc_path_v2.parent
+
+
+def _snapshot(root: Path) -> set[str]:
+    return {path.name for path in root.glob("ht-notify-*")}
+
+
+def test_session_cleanup_removes_directory_left_by_closed_store(root: Path, tmp_path: Path) -> None:
+    before = _snapshot(root)
+    store = _started(tmp_path / "state", tmp_path / "objects")
+    directory = _directory(store)
+    assert directory.parent == root
+    store.close_v2_notifications()
+    store._db.close()
+    assert directory.is_dir()  # production keeps the empty directory
+    remove_new_empty_notify_dirs(root, before)
+    assert not directory.exists()
+
+
+def test_session_cleanup_keeps_directory_with_live_socket(root: Path, tmp_path: Path) -> None:
+    before = _snapshot(root)
+    first, second = (_started(tmp_path / "state", tmp_path / f"objects-{i}") for i in range(2))
+    first.close_v2_notifications()
+    remove_new_empty_notify_dirs(root, before)
+    assert second._ipc_path_v2 is not None and second._ipc_path_v2.is_socket()
+    second.close_v2_notifications()
+    for store in (first, second):
+        store._db.close()
+
+
+def test_session_cleanup_keeps_directory_that_predates_snapshot(root: Path) -> None:
+    existing = root / "ht-notify-preexisting"
+    existing.mkdir(mode=0o700)
+    remove_new_empty_notify_dirs(root, _snapshot(root))
+    assert existing.is_dir()
+
+
+def test_shared_directory_survives_concurrent_start_close(
+    root: Path, tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    # Regression: removing the shared directory on close raced a peer's mkdir/bind
+    # (FileNotFoundError at bind, FileExistsError at mkdir).
+    class SlowBind(socket.socket):
+        def bind(self, address: object) -> None:
+            time.sleep(0.001)  # widen the mkdir-to-bind window a peer's close could hit
+            super().bind(address)  # type: ignore[arg-type]
+
+    monkeypatch.setattr(socket, "socket", SlowBind)
+    threads, cycles = 8, 100
+    stores = [_store(tmp_path / "state", tmp_path / f"objects-{i}") for i in range(threads)]
+    gate = threading.Barrier(threads)
+    errors: list[OSError] = []
+
+    def cycle(store: ChallengeStore) -> None:
+        gate.wait()
+        for _ in range(cycles):
+            try:
+                store._start_notifications_v2()
+                store.close_v2_notifications()
+            except OSError as error:  # keep cycling: one lost race must not end the probe
+                errors.append(error)
+
+    workers = [threading.Thread(target=cycle, args=(store,)) for store in stores]
+    for worker in workers:
+        worker.start()
+    for worker in workers:
+        worker.join()
+    for store in stores:
+        store.close_v2_notifications()
+        store._db.close()
+    assert errors == []
diff --git a/tests/conftest.py b/tests/conftest.py
index f5da5a2..8968ce5 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -3,12 +3,15 @@
 import json
 import os
 import subprocess
+import tempfile
 import uuid
+from collections.abc import Iterator
 from pathlib import Path
 
 import numpy as np
 import pytest
 
+from notify_cleanup import remove_new_empty_notify_dirs
 from systemd_probe import classify_systemd_probe
 
 SEQ_LEN = 1024
@@ -156,3 +159,20 @@ def systemd_manager() -> None:
     if verdict == "error":
         pytest.fail(f"systemd-run probe failed: {reason}")
     pytest.skip(f"system manager transient units unavailable: {reason}")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def _tmp_leak_guard(tmp_path_factory: pytest.TempPathFactory) -> Iterator[None]:
+    """Keep the session's IPC and torchrun leftovers out of the shared temp root.
+
+    torchrun reads PET_LOG_DIR as --log-dir, so its logs land under basetemp instead of a fresh
+    tempdir()/torchelastic_*; child processes inherit os.environ. Nothing outside basetemp is
+    deleted except empty ht-notify-* directories that appeared during the session.
+    """
+    root = Path(tempfile.gettempdir())
+    before = {path.name for path in root.glob("ht-notify-*")}
+    patch = pytest.MonkeyPatch()
+    patch.setenv("PET_LOG_DIR", str(tmp_path_factory.mktemp("torchelastic")))
+    yield
+    patch.undo()
+    remove_new_empty_notify_dirs(root, before)
diff --git a/tests/notify_cleanup.py b/tests/notify_cleanup.py
new file mode 100644
index 0000000..62c0730
--- /dev/null
+++ b/tests/notify_cleanup.py
@@ -0,0 +1,16 @@
+"""Session-end cleanup of the empty ht-notify-* directories that tests leave in the shared /tmp."""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+
+def remove_new_empty_notify_dirs(root: Path, before: set[str]) -> None:
+    """rmdir only: a non-empty directory may be a concurrent same-uid session's live store."""
+    for path in root.glob("ht-notify-*"):
+        if path.name not in before and path.lstat().st_uid == os.getuid():
+            try:
+                path.rmdir()
+            except OSError:
+                continue  # still in use, or a store a test never closed: leave it visible