Address PR feedback: revert Core version, document backpressure, fix test cleanup

- Revert foundry-local-core to 1.0.0rc1 (no dev builds on main)
- Re-add try/except for execute_command_with_binary (1.0.0rc1 lacks symbol)
- Document blocking backpressure behavior in append()
- Remove unnecessary time.sleep(0.5) in error test
- Add stop() cleanup in error test (required for non-daemon thread)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: ruiren_microsoft

sdk/python/requirements-winml.txt

@@ -2,6 +2,6 @@ pydantic>=2.0.0
 requests>=2.32.4
 openai>=2.24.0
 # WinML native binary packages from the ORT-Nightly PyPI feed.
-foundry-local-core-winml==1.0.0.dev202604082039
+foundry-local-core-winml==1.0.0rc1
 onnxruntime-core==1.23.2.3
 onnxruntime-genai-core==0.13.1

sdk/python/requirements.txt

@@ -2,7 +2,7 @@ pydantic>=2.0.0
 requests>=2.32.4
 openai>=2.24.0
 # Standard native binary packages from the ORT-Nightly PyPI feed.
-foundry-local-core==1.0.0.dev202604082039
+foundry-local-core==1.0.0rc1
 onnxruntime-core==1.24.4; sys_platform != "linux"
 onnxruntime-gpu==1.24.4; sys_platform == "linux"
 onnxruntime-genai-core==0.13.1; sys_platform != "linux"

sdk/python/src/detail/core_interop.py

@@ -190,9 +190,15 @@ def _initialize_native_libraries() -> 'NativeBinaryPaths':
                                                       ctypes.c_void_p]  # user_data
         lib.execute_command_with_callback.restype = None
 
-        lib.execute_command_with_binary.argtypes = [ctypes.POINTER(StreamingRequestBuffer),
-                                                     ctypes.POINTER(ResponseBuffer)]
-        lib.execute_command_with_binary.restype = None
+        # execute_command_with_binary is required for live audio streaming.
+        # Guard with try/except until Core packages with this symbol are released.
+        try:
+            lib.execute_command_with_binary.argtypes = [ctypes.POINTER(StreamingRequestBuffer),
+                                                         ctypes.POINTER(ResponseBuffer)]
+            lib.execute_command_with_binary.restype = None
+        except AttributeError:
+            logger.debug("execute_command_with_binary not exported by Core — "
+                         "live audio streaming will not be available until Core is updated")
 
         return paths
 

sdk/python/src/openai/live_audio_transcription_client.py

@@ -168,6 +168,9 @@ def append(self, pcm_data: bytes) -> None:
                     "No active streaming session. Call start() first."
                 )
 
+            # put() blocks if the queue is full (backpressure). This prevents
+            # unbounded memory growth when the native core is slower than
+            # real-time. Capacity is configurable via push_queue_capacity.
             push_queue.put(data_copy)
 
     def get_transcription_stream(self) -> Generator[LiveAudioTranscriptionResponse, None, None]:

sdk/python/test/openai/test_live_audio_transcription.py

@@ -316,14 +316,13 @@ def test_push_error_surfaces_as_exception(self):
 
         session.append(b'\x00' * 3200)
 
-        # Give push loop time to process
-        import time
-        time.sleep(0.5)
-
         with pytest.raises(FoundryLocalException, match="Push failed"):
             for _ in session.get_transcription_stream():
                 pass
 
+        # Cleanup: stop to join the push thread
+        session.stop()
+
     def test_context_manager_calls_stop(self):
         """Verify context manager calls stop on exit."""
         mock_interop = MagicMock(spec=CoreInterop)