diff --git a/tests/test_smoke.py b/tests/test_smoke.py
index a07c4e4..c8b2a7e 100644
--- a/tests/test_smoke.py
+++ b/tests/test_smoke.py
@@ -4182,7 +4182,6 @@ def test_xwayland_tab_renders_and_cycles(xwayland_session):
             f"X11 tab rendered black at ({cx},{cy})"
 
 
-@pytest.mark.flaky
 def test_xwayland_alt_delete_closes_x11_tab(xwayland_session):
     """Alt+Delete on the active X11 tab sends WM_DELETE_WINDOW; xmessage
     honors it, exits, and the escalation never needs to fire."""
@@ -4194,18 +4193,18 @@ def test_xwayland_alt_delete_closes_x11_tab(xwayland_session):
         s.wait_active(1)
     got = s.cycle_ack(+1)
     assert got == 0
-    d = dict((int(a), int(p)) for a, p in CLAIM_RE.findall(s.read_log()))
-    xpid = d[0]
     s.chord_ack(ALT, DELETE)
-    since = time.monotonic()
-    while time.monotonic() - since < 15.0:
-        try:
-            os.kill(xpid, 0)
-        except ProcessLookupError:
-            break  # xmessage exited from the graceful close
-        time.sleep(0.1)
-    else:
-        pytest.fail(f"xmessage (pid {xpid}) survived Alt+Delete")
+    # Event, not wall-clock: kick's SIGCHLD handler logs every reaped child
+    # unbuffered.  Watch the LAUNCHER pid ([launch]), not the claim pid: the
+    # claim is the X client's pid, but sh -c forks so the reaped direct
+    # child (which exits once its client does) is a different pid.  The 30s
+    # backstop bounds hangs only.  (Whether xmessage exits gracefully via
+    # WM_DELETE or via the escalation SIGTERM, the launcher reaps the same;
+    # this test asserts teardown + survival either way.)
+    m = re.search(r"\[launch\] app=0 pid=(\d+)", s.read_log())
+    assert m, "xmessage launcher never logged"
+    # NOTE: [child] uses `pid ` (space), unlike [launch]/[claim] `pid=`.
+    s.wait_for(r"\[child\] pid %s exited" % m.group(1), timeout=30.0)
     # The compositor stays up with the remaining Wayland tab.
     assert s.proc.poll() is None, "compositor exited after closing X11 tab"
     assert s.active() == 1