feat(auto): выбран ONNX по умолчанию без CUDA

- Зачем:
  - пользователям без NVIDIA нужен самый быстрый и читаемый CPU-профиль без дополнительных параметров.
- Что:
  - auto-политика изменена на CUDA при наличии nvidia-smi, иначе ONNX GigaAM RNN-T int8.
  - сохранён приоритет явных значений CLI и конфигурации для OpenVINO и FasterWhisper CPU.
  - обновлены тесты, README, PRD, ADR, GPU-документация и вывод benchmark.
- Проверка:
  - uv run pytest -q: 232 passed, 1 skipped.
  - uv lock --check и git diff --cached --check.

tests/test_cli.py

@@ -73,27 +73,32 @@ def test_cli_happy_path_exit_code_zero(tmp_path):
 def test_cli_default_options_passed_to_transcribe(tmp_path):
     audio = tmp_path / "test.mp3"
     audio.write_bytes(b"fake")
-    result = _make_result()
+    result = _make_result(device_used="onnx")
     model = _make_model()
     backend = _make_backend()
-    tfr = _make_tfr(result=result, model=model, backend=backend)
+    tfr = _make_tfr(result=result, model=model, actual_device="onnx", backend=backend)
     mock_transcribe_file = MagicMock(return_value=tfr)
 
     with (
         patch("local_transcriber.cli.load_config", return_value={}),
         patch("local_transcriber.cli.validate_input_file", return_value=audio),
-        patch("local_transcriber.cli.detect_device", return_value="cpu"),
-        patch("local_transcriber.cli.load_model", return_value=(model, "cpu", backend, "/models/medium")),
+        patch("local_transcriber.cli.detect_device", return_value="onnx"),
+        patch(
+            "local_transcriber.cli.load_model",
+            return_value=(model, "onnx", backend, "/models/gigaam-v3-e2e-rnnt"),
+        ),
         patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
         patch("local_transcriber.cli.write_transcript"),
     ):
-        runner.invoke(app, [str(audio)])
+        out = runner.invoke(app, [str(audio)])
 
     call_kwargs = mock_transcribe_file.call_args[1]
-    assert call_kwargs["model_name"] == "medium"
-    assert call_kwargs["compute_type"] == "float32"
+    assert call_kwargs["model_name"] == "gigaam-v3-e2e-rnnt"
+    assert call_kwargs["compute_type"] == "int8"
     assert call_kwargs["language"] == "ru"
     assert call_kwargs["on_segment"] is None  # verbose=False
+    assert "Модель: gigaam-v3-e2e-rnnt" in out.output
+    assert "Устройство: onnx" in out.output
 
 
 def test_cli_custom_options(tmp_path):
@@ -660,6 +665,45 @@ def test_cli_config_applied(tmp_path):
     assert mock_load_model.call_args[0][0] == "tiny"
 
 
+def test_cli_config_overrides_auto_device(tmp_path):
+    audio = tmp_path / "test.mp3"
+    audio.write_bytes(b"fake")
+    model = _make_model()
+    backend = _make_backend()
+    result = _make_result(device_used="openvino-cpu")
+    tfr = _make_tfr(
+        result=result,
+        model=model,
+        actual_device="openvino-cpu",
+        backend=backend,
+    )
+    mock_detect_device = MagicMock(return_value="openvino-cpu")
+    mock_load_model = MagicMock(
+        return_value=(model, "openvino-cpu", backend, "/models/medium")
+    )
+
+    with (
+        patch(
+            "local_transcriber.cli.load_config",
+            return_value={
+                "device": "openvino-cpu",
+                "model": "medium",
+                "compute_type": "int8",
+            },
+        ),
+        patch("local_transcriber.cli.validate_input_file", return_value=audio),
+        patch("local_transcriber.cli.detect_device", mock_detect_device),
+        patch("local_transcriber.cli.load_model", mock_load_model),
+        patch("local_transcriber.cli._transcribe_file", return_value=tfr),
+        patch("local_transcriber.cli.write_transcript"),
+    ):
+        out = runner.invoke(app, [str(audio)])
+
+    assert out.exit_code == 0
+    assert mock_detect_device.call_args_list[0].args == ("openvino-cpu",)
+    assert mock_load_model.call_args.args[:3] == ("medium", "openvino-cpu", "int8")
+
+
 def test_cli_cli_overrides_config(tmp_path):
     audio = tmp_path / "test.mp3"
     audio.write_bytes(b"fake")

tests/test_config.py

@@ -71,11 +71,17 @@ def test_load_config_invalid_device(tmp_path):
 
 
 def test_resolve_defaults_cli_wins():
-    config = {"model": "tiny", "language": "en"}
-    cli = {"model": "small", "language": None, "device": None, "compute_type": None}
+    config = {"model": "tiny", "language": "en", "device": "openvino-cpu"}
+    cli = {
+        "model": "small",
+        "language": None,
+        "device": "onnx",
+        "compute_type": None,
+    }
     result = resolve_defaults(cli, config)
     assert result["model"] == "small"
     assert result["language"] == "en"
+    assert result["device"] == "onnx"
 
 
 def test_resolve_defaults_config_wins():

tests/test_utils.py

@@ -61,27 +61,28 @@ def test_detect_device_explicit_passthrough():
     """Явные device strings проходят без изменений."""
     assert detect_device("cpu") == "cpu"
     assert detect_device("cuda") == "cuda"
+    assert detect_device("onnx") == "onnx"
     assert detect_device("openvino-gpu") == "openvino-gpu"
     assert detect_device("openvino-cpu") == "openvino-cpu"
 
 
-def test_detect_device_auto_openvino_gpu():
-    """auto + нет nvidia-smi + есть OpenVINO GPU → openvino-gpu."""
+def test_detect_device_auto_onnx_even_with_openvino_gpu():
+    """auto + нет nvidia-smi → onnx, даже если доступен OpenVINO GPU."""
     with (
         patch("local_transcriber.utils.shutil.which", return_value=None),
         patch("local_transcriber.utils._is_openvino_gpu_available", return_value=True),
     ):
-        assert detect_device("auto") == "openvino-gpu"
+        assert detect_device("auto") == "onnx"
 
 
-def test_detect_device_auto_openvino_cpu():
-    """auto + нет nvidia-smi + есть OpenVINO, нет GPU → openvino-cpu."""
+def test_detect_device_auto_onnx_even_with_openvino_cpu():
+    """auto + нет nvidia-smi → onnx, даже если доступен OpenVINO CPU."""
     with (
         patch("local_transcriber.utils.shutil.which", return_value=None),
         patch("local_transcriber.utils._is_openvino_gpu_available", return_value=False),
         patch("local_transcriber.utils._is_openvino_available", return_value=True),
     ):
-        assert detect_device("auto") == "openvino-cpu"
+        assert detect_device("auto") == "onnx"
 
 
 def test_detect_device_cuda_over_openvino():
@@ -93,14 +94,14 @@ def test_detect_device_cuda_over_openvino():
         assert detect_device("auto") == "cuda"
 
 
-def test_detect_device_auto_cpu_fallback():
-    """Ни nvidia-smi, ни openvino → cpu."""
+def test_detect_device_auto_onnx_without_accelerators():
+    """Без CUDA auto выбирает ONNX CPU."""
     with (
         patch("local_transcriber.utils.shutil.which", return_value=None),
         patch("local_transcriber.utils._is_openvino_gpu_available", return_value=False),
         patch("local_transcriber.utils._is_openvino_available", return_value=False),
     ):
-        assert detect_device("auto") == "cpu"
+        assert detect_device("auto") == "onnx"
 
 
 def test_detect_device_openvino_resolves_to_gpu():