Refactor tool-speechtotext: extract sttlib shared library and add tests

Extract duplicated code (Whisper loading, audio recording, transcription,
VAD processing) into reusable sttlib/ package. Rewrite all 3 scripts as
thin wrappers. Add 24 unit tests with mocked hardware. Fix GPU fallback
bug in assistant.py and args.system assignment bug.

python/tool-speechtotext/tests/test_whisper_loader.py

@@ -0,0 +1,37 @@
+from unittest.mock import patch, MagicMock
+from sttlib.whisper_loader import load_whisper_model
+
+
+@patch("sttlib.whisper_loader.WhisperModel")
+def test_gpu_success(mock_cls):
+    mock_model = MagicMock()
+    mock_cls.return_value = mock_model
+
+    result = load_whisper_model("base")
+
+    assert result is mock_model
+    mock_cls.assert_called_once_with("base", device="cuda", compute_type="float16")
+
+
+@patch("sttlib.whisper_loader.WhisperModel")
+def test_gpu_fails_cpu_fallback(mock_cls):
+    mock_model = MagicMock()
+    mock_cls.side_effect = [RuntimeError("no CUDA"), mock_model]
+
+    result = load_whisper_model("base")
+
+    assert result is mock_model
+    assert mock_cls.call_count == 2
+    _, kwargs = mock_cls.call_args
+    assert kwargs == {"device": "cpu", "compute_type": "int8"}
+
+
+@patch("sttlib.whisper_loader.WhisperModel")
+def test_both_fail_propagates(mock_cls):
+    mock_cls.side_effect = RuntimeError("no device")
+
+    try:
+        load_whisper_model("base")
+        assert False, "Should have raised"
+    except RuntimeError:
+        pass