[Bugfix][MiniCPM-o] Fix async-chunk snapshot replacement and prompt cleanup (#6406)

Signed-off-by: natureofnature <wzliu@connect.hku.hk>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>

tests/model_executor/models/minicpmo_4_5/test_code2wav_batching.py

@@ -537,6 +537,15 @@ def test_empty_final_ignores_generation_scheduler_placeholder_token():
         # The runner injects the engine request id on every step (GPU
         # _preprocess, NPU _gather_runtime_additional_information)...
         {"request_id": "a", "meta": {"request_id": "a"}},
+        # Runtime snapshot bookkeeping alone is not a Talker payload.
+        {
+            "request_id": "a",
+            "meta": {
+                "request_id": "a",
+                "num_processed_tokens": 0,
+                "resumable": True,
+            },
+        },
         # ...but a pre-warm step can also reach the model with nothing at all.
         {},
     ],

tests/model_executor/stage_input_processors/test_minicpmo_4_5_async_chunk.py

@@ -423,6 +423,12 @@ def test_duplex_turn_end_closes_epoch_and_next_turn_restarts_sequence() -> None:
         request,
         True,
     )
+    duplicate_boundary = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(turn_id=7, turn_end=True),
+        request,
+        True,
+    )
     next_turn = tts2code2wav_async_chunk(
         manager,
         _duplex_delta(20, turn_id=8),
@@ -435,6 +441,10 @@ def test_duplex_turn_end_closes_epoch_and_next_turn_restarts_sequence() -> None:
     assert turn_end.meta.last_chunk is True
     assert turn_end.meta.turn_end is True
     assert turn_end.meta.is_segment_finished.item() is True
+    assert turn_end.meta.replace_runtime_additional_information is True
+    assert duplicate_boundary is not None
+    assert duplicate_boundary.meta.is_segment_finished.item() is True
+    assert duplicate_boundary.meta.replace_runtime_additional_information is True
     assert next_turn.meta.cache_epoch == turn_end.meta.cache_epoch + 1
     assert next_turn.meta.chunk_seq == 0
     assert next_turn.meta.last_chunk is False