diff --git a/pyproject.toml b/pyproject.toml
index 0a23dce5..d06c0f5e 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -84,6 +84,11 @@ testing = [
   # interactions stay predictable across dev installs).
   "pytest>=9.0.3",  # CVE-2025-71176 (insecure tmpdir) patched in 9.0.3
   "pexpect>=4.9.0,<5",
+  # per-test wall-clock bound (used via
+  # `@pytest.mark.timeout(..., method='thread')` on the
+  # known-hanging `subint`-backend audit tests; see
+  # `ai/conc-anal/subint_*_issue.md`).
+  "pytest-timeout>=2.3",
 ]
 repl = [
   "pyperclip>=1.9.0",
diff --git a/tests/discovery/test_registrar.py b/tests/discovery/test_registrar.py
index 6f34b117..bd015608 100644
--- a/tests/discovery/test_registrar.py
+++ b/tests/discovery/test_registrar.py
@@ -517,6 +517,22 @@ async def kill_transport(
 
 
 # @pytest.mark.parametrize('use_signal', [False, True])
+#
+# Wall-clock bound via `pytest-timeout` (`method='thread'`).
+# Under `--spawn-backend=subint` this test can wedge in an
+# un-Ctrl-C-able state (abandoned-subint + shared-GIL
+# starvation → signal-wakeup-fd pipe fills → SIGINT silently
+# dropped; see `ai/conc-anal/subint_sigint_starvation_issue.md`).
+# `method='thread'` is specifically required because `signal`-
+# method SIGALRM suffers the same GIL-starvation path and
+# wouldn't fire the Python-level handler.
+# At timeout the plugin hard-kills the pytest process — that's
+# the intended behavior here; the alternative is an unattended
+# suite run that never returns.
+@pytest.mark.timeout(
+    3,  # NOTE should be a 2.1s happy path.
+    method='thread',
+)
 def test_stale_entry_is_deleted(
     debug_mode: bool,
     daemon: subprocess.Popen,