diff --git a/custom_components/nvidia_shield/connection.py b/custom_components/nvidia_shield/connection.py
index 3ac1162..5b4769b 100644
--- a/custom_components/nvidia_shield/connection.py
+++ b/custom_components/nvidia_shield/connection.py
@@ -276,7 +276,7 @@ class ShieldConnectionManager:
 
         try:
             if not self.connected and not await self._poller.connect():
-                self._note_failure(backoff)
+                await self._note_failure(backoff)
                 return self._jitter(backoff)
             if self._failures:
                 self._failures = 0
@@ -285,11 +285,11 @@ class ShieldConnectionManager:
             await self._refresh()
             return ADB_POLL_INTERVAL
         except Exception as err:
-            self._note_failure(backoff)
+            await self._note_failure(backoff)
             _LOGGER.debug("Watchdog cycle failed: %s", err)
             return self._jitter(backoff)
 
-    def _note_failure(self, backoff: float) -> None:
+    async def _note_failure(self, backoff: float) -> None:
         """Bookkeep a failed cycle: log sparingly, drop stale store, notify."""
         self._failures += 1
         if self._failures == 1 or self._failures % 6 == 0:
@@ -300,6 +300,13 @@ class ShieldConnectionManager:
             )
         if not self._down_since:
             self._down_since = True
+            # Reconnect decision (health_check's documented role): close the
+            # socket so the next cycle rebuilds it from scratch instead of
+            # trusting shell()'s auto-reconnect on a broken transport.
+            try:
+                await self._poller.disconnect()
+            except Exception:
+                _LOGGER.debug("disconnect during reconnect decision failed", exc_info=True)
             self.store.clear(*ALL_FIELDS)
             self._notify()
 
diff --git a/tests/test_connection.py b/tests/test_connection.py
index 1a04042..d59034f 100644
--- a/tests/test_connection.py
+++ b/tests/test_connection.py
@@ -219,8 +219,8 @@ class TestReconnectBackoff:
         # 5→10→20→40→80→160→300 growth with easing; the last must not exceed
         # the cap (with jitter).
         assert max(delays) <= 300.0 * 1.25
-        assert delays[1] > delays[0] * 1.5
-        assert delays[-1] < 300.0 * 1.25
+        assert delays[1] >= delays[0] * 1.5
+        assert delays[-1] <= 300.0 * 1.25
 
     async def test_repeated_failures_do_not_spam_listeners(self):
         poller = FakePoller()
@@ -245,6 +245,21 @@ class TestReconnectBackoff:
         assert manager._failures == 0
         assert await manager._watchdog_cycle(5.0) == ADB_POLL_INTERVAL
 
+    async def test_failed_wake_poll_forces_reconnect(self):
+        """health_check drives the reconnect decision: a cycle where is_awake
+        fails closes the socket so the next cycle reconnects from scratch."""
+        poller = FakePoller()
+        poller.connected = True
+        poller.awake = None  # wake poll fails while the flag says "connected"
+        manager = _manager(poller)
+
+        delay = await manager._watchdog_cycle(5.0)
+
+        assert 5.0 * 0.75 <= delay <= 5.0 * 1.25  # F5-style symmetric bound
+        assert poller.disconnect_calls == 1
+        assert poller.connected is False
+        assert manager._failures == 1
+
     async def test_connect_failure_clears_stale_store_state(self):
         poller = FakePoller()
         poller.connect_failures = 1