fix: smart chunking, extraction, and process-safety issuesHEADmain

1 files changed, 31 insertions, 6 deletions

diff --git a/app/tests/test_backends_servers.py b/app/tests/test_backends_servers.py
index 9fd71f2..283df95 100644
--- a/app/tests/test_backends_servers.py
+++ b/app/tests/test_backends_servers.py
@@ -370,19 +370,44 @@ class ReapTests(unittest.TestCase):
 
 
 class KillPidTests(unittest.TestCase):
-    """_kill_pid: the reap check ends the grace wait before SIGKILL."""
+    """_kill_pid: the group probe decides the wait's end, not the reap.
 
-    def test_reaped_child_ends_wait_without_sigkill(self):
+    Reaping the launcher proves nothing about the rest of its process
+    group: workers can outlive it and must still get the SIGKILL
+    escalation.
+    """
+
+    def test_empty_group_ends_wait_without_sigkill(self):
+        # Single-process server: the launcher is reaped and the group is
+        # empty (killpg(0) raises) — success immediately, no SIGKILL.
         with patch("os.getpgid", return_value=4242), \
-                patch("os.killpg") as mk_killpg, \
-                patch("os.waitpid", return_value=(4242, 0)) as mk_waitpid, \
+                patch("os.killpg",
+                      side_effect=[None, ProcessLookupError]) as mk_killpg, \
+                patch("os.waitpid", return_value=(4242, 0)), \
                 patch("time.sleep") as mk_sleep:
             ok = servers._kill_pid(4242)
         self.assertTrue(ok)
-        mk_killpg.assert_called_once_with(4242, signal.SIGTERM)
-        mk_waitpid.assert_called_once_with(4242, os.WNOHANG)
+        self.assertEqual(mk_killpg.call_args_list[0].args,
+                         (4242, signal.SIGTERM))
+        self.assertEqual(mk_killpg.call_args_list[-1].args, (4242, 0))
         mk_sleep.assert_not_called()
 
+    def test_reaped_leader_with_live_workers_escalates_to_sigkill(self):
+        # A launcher that dies on SIGTERM while its workers ignore it:
+        # reaping the leader must not end the stop, the surviving group
+        # keeps burning the grace period and then gets SIGKILLed.
+        with patch("os.getpgid", return_value=4242), \
+                patch("os.killpg", return_value=None) as mk_killpg, \
+                patch("os.waitpid", return_value=(4242, 0)), \
+                patch("time.sleep"):
+            ok = servers._kill_pid(4242)
+        self.assertTrue(ok)
+        calls = mk_killpg.call_args_list
+        self.assertEqual(calls[0].args, (4242, signal.SIGTERM))
+        self.assertEqual(calls[-1].args, (4242, signal.SIGKILL))
+        # The group was probed repeatedly while waiting for the workers.
+        self.assertGreater(len(calls), 2)
+
     def test_escalates_to_sigkill_when_child_stays_alive(self):
         with patch("os.getpgid", return_value=4242), \
                 patch("os.killpg") as mk_killpg, \