diff --git a/tests/ut/cpp/hierarchical/test_scheduler.cpp b/tests/ut/cpp/hierarchical/test_scheduler.cpp
index f0036c5ce..7ad8f8c89 100644
--- a/tests/ut/cpp/hierarchical/test_scheduler.cpp
+++ b/tests/ut/cpp/hierarchical/test_scheduler.cpp
@@ -69,6 +69,7 @@ struct MockMailboxWorker {
     std::mutex run_mu;
     std::condition_variable run_cv;
     std::atomic<bool> should_complete{false};
+    std::atomic<bool> drain_mode{false};
     int32_t next_error_code{0};
     std::string next_error_msg;
     std::atomic<bool> is_running{false};
@@ -116,6 +117,15 @@ struct MockMailboxWorker {
         run_cv.notify_one();
     }
 
+    // Persistent teardown mode: every dispatch — including one arriving after
+    // this call — completes itself, so Scheduler::stop() can always join.
+    void drain() {
+        drain_mode.store(true, std::memory_order_release);
+        std::lock_guard<std::mutex> lk(run_mu);
+        should_complete.store(true, std::memory_order_release);
+        run_cv.notify_one();
+    }
+
     // The child publishes acceptance into the sticky word, not the state.
     void write_task_accepted() {
         auto *ptr = reinterpret_cast<int32_t *>(static_cast<char *>(mailbox_ptr()) + MAILBOX_OFF_ACCEPTED);
@@ -176,7 +186,8 @@ struct MockMailboxWorker {
                 {
                     std::unique_lock<std::mutex> lk(run_mu);
                     run_cv.wait(lk, [this] {
-                        return should_complete.load(std::memory_order_acquire);
+                        return should_complete.load(std::memory_order_acquire) ||
+                               drain_mode.load(std::memory_order_acquire);
                     });
                     should_complete.store(false, std::memory_order_relaxed);
                 }
@@ -502,6 +513,7 @@ struct SchedulerFixture : public ::testing::Test {
     }
 
     void TearDown() override {
+        mock_worker.drain();
         sched.stop();
         manager.stop();
         allocator.shutdown();
@@ -916,6 +928,9 @@ struct GroupSchedulerFixture : public ::testing::Test {
     }
 
     void TearDown() override {
+        worker_a.drain();
+        worker_b.drain();
+        worker_c.drain();
         sched.stop();
         manager.stop();
         allocator.shutdown();
@@ -1259,6 +1274,30 @@ TEST_F(GroupSchedulerFixture, ConsecutiveGroupsReserveOnlyBlockedHeadTargets) {
     wait_consumed(single_c.task_slot);
 }
 
+TEST_F(GroupSchedulerFixture, TearDownDrainsCurrentAndQueuedDispatches) {
+    // The verification is the teardown itself: work is left deliberately in
+    // both states — running and queued-but-undispatched — so teardown drains
+    // a worker mid-task and one whose dispatch has not happened yet.
+    {
+        std::lock_guard<std::mutex> scheduler_pause(sched.loop_mutex());
+        (void)orch.submit_next_level_group(
+            C(84), {single_tensor_args(0x106, TensorArgType::OUTPUT), single_tensor_args(0x107, TensorArgType::OUTPUT)},
+            cfg, {0, 1}
+        );
+        (void)orch.submit_next_level_group(
+            C(85), {single_tensor_args(0x108, TensorArgType::OUTPUT), single_tensor_args(0x109, TensorArgType::OUTPUT)},
+            cfg, {1, 2}
+        );
+        (void)orch.submit_next_level(C(86), single_tensor_args(0x10A, TensorArgType::OUTPUT), cfg, 0);
+        (void)orch.submit_next_level(C(87), single_tensor_args(0x10B, TensorArgType::OUTPUT), cfg, 2);
+    }
+
+    worker_a.wait_running();
+    worker_b.wait_running();
+    EXPECT_TRUE(worker_a.is_running.load(std::memory_order_acquire));
+    EXPECT_TRUE(worker_b.is_running.load(std::memory_order_acquire));
+}
+
 TEST_F(GroupSchedulerFixture, LaunchableGroupPrecedesConflictingSingles) {
     auto running_a = orch.submit_next_level(C(73), single_tensor_args(0xF4, TensorArgType::OUTPUT), cfg, 0);
     auto running_b = orch.submit_next_level(C(74), single_tensor_args(0xF5, TensorArgType::OUTPUT), cfg, 1);
@@ -1743,6 +1782,8 @@ struct MixedTypeSchedulerFixture : public ::testing::Test {
     }
 
     void TearDown() override {
+        next_level_worker.drain();
+        sub_worker.drain();
         sched.stop();
         manager.stop();
         allocator.shutdown();