Close leaked listening socket on update restart

Werkzeug sets its server socket inheritable unconditionally, so the
os.execv restart carried it into the new process as a zombie listener:
one PID listening on both 5100 (never accepted) and 5101 (the real
server). Mark all fds above stderr close-on-exec before exec'ing so
the old socket dies and the new server rebinds the original port.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CHANGELOG.md

@@ -9,6 +9,10 @@ Version numbers follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html
 
 ## [Unreleased]
 
+### Fixed
+
+- **Update restart leaked the listening socket and hopped to port 5101** — Werkzeug marks its server socket inheritable (`srv.socket.set_inheritable(True)`, unconditionally, for its debug reloader), so the in-app update's `os.execv` restart carried the old listening socket into the new process as a zombie listener: same PID listening on both 5100 (never accepted — clients hang) and 5101 (the actual server). The 1.7.3 `SO_REUSEADDR`/grace-period fix couldn't help because the port genuinely was occupied — by the restarting process itself. `_restart_self()` now marks every fd above stderr close-on-exec before the exec (`_mark_fds_cloexec()`, enumerating `/proc/self/fd` on Linux), so the old socket dies with the exec and the new server rebinds 5100 immediately.
+
 ---
 
 ## [1.7.5] — 2026-06-11

routes/updates.py

@@ -118,13 +118,34 @@ def apply_update() -> dict:
             "from": chk["current"], "to": chk["latest"]}
 
 
+def _mark_fds_cloexec() -> None:
+    """Mark every fd above stderr close-on-exec.
+
+    Werkzeug calls ``srv.socket.set_inheritable(True)`` unconditionally
+    (for its debug reloader), so without this the listening socket leaks
+    into the exec'd process: it sits on the port as a zombie listener no
+    one accepts from, the port probe sees the port as busy, and the new
+    server hops to port+1 while clients hang against the dead socket.
+    """
+    try:
+        fds = [int(f) for f in os.listdir("/proc/self/fd")]   # Linux
+    except (OSError, ValueError):
+        fds = list(range(3, 4096))
+    for fd in fds:
+        if fd > 2:
+            try:
+                os.set_inheritable(fd, False)
+            except OSError:
+                pass
+
+
 def _restart_self() -> None:
     """Re-exec the current process so the updated code is loaded.
 
     Keeps the same PID, so it works both under systemd and when launched
-    manually via start_gdpr.sh. Listening sockets are close-on-exec, so
-    the new process can rebind the port.
+    manually via start_gdpr.sh.
     """
+    _mark_fds_cloexec()
     try:
         os.execv(sys.executable, [sys.executable] + sys.argv)
     except OSError:

tests/test_updates.py

@@ -188,6 +188,23 @@ def test_apply_installs_requirements_when_changed(client, monkeypatch):
     assert "pip" in pip_calls[0] and "-r" in pip_calls[0]
 
 
+# ── Restart fd hygiene ────────────────────────────────────────────────────────
+
+def test_mark_fds_cloexec_unmarks_inheritable_socket():
+    """Werkzeug sets the listening socket inheritable; the restart must undo
+    that or the socket leaks through execv and squats on the port."""
+    import socket
+    import routes.updates as upd
+    s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    try:
+        s.set_inheritable(True)
+        assert s.get_inheritable() is True
+        upd._mark_fds_cloexec()
+        assert s.get_inheritable() is False
+    finally:
+        s.close()
+
+
 # ── /api/update/settings ──────────────────────────────────────────────────────
 
 def test_settings_roundtrip(client, monkeypatch):