test(server): fix flaky testSynchronousExecutorDoesNotLeakChainEntries

server-common/src/test/java/org/a2aproject/sdk/server/events/MainEventBusProcessorPushNotificationOrderTest.java

@@ -146,8 +146,17 @@ public void testSynchronousExecutorDoesNotLeakChainEntries() throws InterruptedE
         // One entry per DISTINCT task leaks, not repeated pushes for the same one --
         // a single task's own map slot just gets overwritten each time. Only several
         // different task IDs reveal growth.
+        //
+        // Each task produces TWO push notifications: one for the initial Task event and one for the
+        // StatusUpdate event (both implement StreamingEventKind). With a synchronous executor,
+        // pushTask runs inside compute()'s lambda -- before the entry is stored in the map -- so
+        // the latch fires before the entry is even visible. To avoid a TOCTOU race between the
+        // while-loop exit (count==0 transiently) and the assertEquals re-reading count, we:
+        //   1. Set the latch to cover ALL push notifications (taskCount * 2).
+        //   2. Always sleep before the first count check so the last cleanup has time to run.
+        //   3. Assert on the captured count, not a fresh read.
         int taskCount = 10;
-        CountDownLatch latch = new CountDownLatch(taskCount);
+        CountDownLatch latch = new CountDownLatch(taskCount * 2);
         PushNotificationSender sender = (event, snapshot) -> latch.countDown();
 
         mainEventBusProcessor = new MainEventBusProcessor(mainEventBus, taskStore, sender, queueManager);
@@ -171,10 +180,16 @@ public void testSynchronousExecutorDoesNotLeakChainEntries() throws InterruptedE
         assertTrue(latch.await(5, TimeUnit.SECONDS), "All push notifications should complete");
 
         long deadline = System.nanoTime() + TimeUnit.SECONDS.toNanos(5);
-        while (mainEventBusProcessor.pushNotificationChainCount() != 0 && System.nanoTime() < deadline) {
+        int count;
+        do {
+            // Always sleep before checking: with the sync executor, pushTask fires inside
+            // compute()'s lambda (before the entry is stored), so the latch may reach 0
+            // slightly before the final whenComplete removal runs. A brief sleep lets the
+            // processor thread finish that last cleanup step.
             Thread.sleep(10);
-        }
-        assertEquals(0, mainEventBusProcessor.pushNotificationChainCount(),
+            count = mainEventBusProcessor.pushNotificationChainCount();
+        } while (count != 0 && System.nanoTime() < deadline);
+        assertEquals(0, count,
                 "A completed task's chain entry must not be left in the map");
     }