fix(perception): rendezvous completed gpu handoff

tests/test_perception_joint_pilot.py

@@ -176,6 +176,44 @@ def test_gpu_stage_preserves_order_and_reserves_bounded_output_before_compute(pi
         assert stage.close()
 
 
+def test_waiting_consumer_uses_direct_handoff_without_evicting_ingress(pilot):
+    mailbox = pilot("pilot_queue").Mailbox(capacity=2, byte_limit=100)
+    compute_started = threading.Event()
+    release_compute = threading.Event()
+
+    def compute(bundle):
+        if bundle["sequence"] == 0:
+            compute_started.set()
+            assert release_compute.wait(1)
+        return bundle["sequence"] * 10
+
+    stage = pilot("pilot_scheduler").GpuStage(mailbox, compute, threading.Event())
+    bundles = [{"sequence": sequence, "payload_bytes": 10} for sequence in range(3)]
+    mailbox.put(bundles[0])
+    assert compute_started.wait(1)
+    mailbox.put(bundles[1])
+    mailbox.put(bundles[2])
+    mailbox.finish()
+    received = []
+    consumer = threading.Thread(target=lambda: received.append(stage.take()))
+    consumer.start()
+    with stage.output_condition:
+        assert stage.output_condition.wait_for(lambda: stage.consumer_waiting, timeout=1)
+    release_compute.set()
+    consumer.join(timeout=1)
+    assert not consumer.is_alive()
+    assert received[0][0]["sequence"] == 0
+    assert not mailbox.dropped
+    assert stage.direct_handoffs == 1
+    for expected in (1, 2):
+        bundle, result = stage.take()
+        assert (bundle["sequence"], result) == (expected, expected * 10)
+    for bundle in bundles:
+        mailbox.release(bundle)
+    assert stage.take() is None
+    assert mailbox.bytes == 0 and stage.close()
+
+
 def test_gpu_stage_propagates_failure_and_stops_waiting_for_input(pilot):
     mailbox = pilot("pilot_queue").Mailbox(capacity=1)