feat(cli): --threads для управления CPU-потоками + бенчмарк CUDA compute_type

- Зачем:
  - CTranslate2 по умолчанию использует 4 потока; на многоядерных CPU (8+ ядер)
    это неоптимально — --threads 8 даёт +13% ускорения.
  - Не было данных по int8_float32/int8_float16 на NVIDIA GPU.

- Что:
  - --threads / -t: новый CLI-флаг, пробрасывается через load_model →
    backend.create_model(cpu_threads=...) → WhisperModel(cpu_threads=...).
  - Валидация min=0 на входе (typer), Backend протокол синхронизирован.
  - docs/gpu.md: результаты бенчмарка 6 комбинаций CUDA compute_type
    (medium/large-v3 × float16/int8_float32/int8_float16) на двух файлах
    (16 мин и 46 мин). Ключевой вывод: float16 — оптимальный дефолт;
    large-v3 ненадёжен на длинных записях.
  - README: --threads добавлен в таблицу опций.
  - Фикс теста: test_resolve_repo_explicit_unsupported_pair_raises обновлён
    под добавление medium fp16 модели.

- Проверка:
  - uv run pytest: 157 passed.
  - transcribe file.mp4 --device cpu --threads 8: 277с vs 320с (дефолт).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

tests/test_backend_openvino.py

@@ -31,7 +31,7 @@ def test_resolve_repo_explicit_unsupported_pair_raises():
     """Явный --compute-type с несуществующей парой → ошибка."""
     backend = OpenVINOBackend(compute_type_explicit=True)
     with pytest.raises(ValueError, match="недоступна с compute_type='fp16'"):
-        backend._resolve_repo("medium", "fp16")
+        backend._resolve_repo("small", "fp16")
 
 
 def test_resolve_repo_explicit_unknown_model_raises():

tests/test_cli.py

@@ -862,3 +862,62 @@ def test_cli_openvino_alias_resolves_to_gpu(tmp_path):
     assert out.exit_code == 0
     # detect_device("openvino") resolved to "openvino-gpu", load_model receives it
     assert mock_load_model.call_args[0][1] == "openvino-gpu"
+
+
+# === --threads ===
+
+
+def test_cli_threads_passed_to_load_model(tmp_path):
+    """--threads передаётся в load_model как cpu_threads."""
+    audio = tmp_path / "test.mp3"
+    audio.write_bytes(b"fake")
+    result = _make_result()
+    model = _make_model()
+    backend = _make_backend()
+    tfr = _make_tfr(result=result, model=model, backend=backend)
+    mock_load_model = MagicMock(return_value=(model, "cpu", backend, "/models/medium"))
+
+    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", mock_load_model),
+        patch("local_transcriber.cli._transcribe_file", return_value=tfr),
+        patch("local_transcriber.cli.write_transcript"),
+    ):
+        out = runner.invoke(app, [str(audio), "--threads", "8"])
+
+    assert out.exit_code == 0
+    assert mock_load_model.call_args.kwargs["cpu_threads"] == 8
+
+
+def test_cli_threads_default_zero(tmp_path):
+    """Без --threads load_model получает cpu_threads=0."""
+    audio = tmp_path / "test.mp3"
+    audio.write_bytes(b"fake")
+    result = _make_result()
+    model = _make_model()
+    backend = _make_backend()
+    tfr = _make_tfr(result=result, model=model, backend=backend)
+    mock_load_model = MagicMock(return_value=(model, "cpu", backend, "/models/medium"))
+
+    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", 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_load_model.call_args.kwargs["cpu_threads"] == 0
+
+
+def test_cli_threads_negative_rejected(tmp_path):
+    """--threads с отрицательным значением отклоняется typer (min=0)."""
+    audio = tmp_path / "test.mp3"
+    audio.write_bytes(b"fake")
+    out = runner.invoke(app, [str(audio), "--threads", "-1"])
+    assert out.exit_code != 0

tests/test_transcriber.py

@@ -321,6 +321,34 @@ def test_load_model_returns_backend_and_path(mock_get_backend):
     assert actual_device == "cpu"
 
 
+@patch("local_transcriber.transcriber.get_backend")
+def test_load_model_passes_cpu_threads_to_backend(mock_get_backend):
+    backend = _make_backend(model_path="/mock/model/path")
+    mock_get_backend.return_value = backend
+
+    load_model("tiny", "cpu", "int8", cpu_threads=8)
+
+    assert backend.create_model.call_args.kwargs["cpu_threads"] == 8
+
+
+@patch("local_transcriber.transcriber.get_backend")
+def test_load_model_fallback_preserves_cpu_threads(mock_get_backend):
+    cuda_backend = _make_backend(create_model_error=RuntimeError("CUDA out of memory"))
+    cpu_model = MagicMock()
+    cpu_backend = _make_backend(model=cpu_model, model_path="/mock/cpu/model")
+
+    def backend_for_device(device, **kwargs):
+        return cuda_backend if device == "cuda" else cpu_backend
+
+    mock_get_backend.side_effect = backend_for_device
+
+    with pytest.warns(UserWarning, match="Переключение на CPU"):
+        load_model("tiny", "cuda", "int8", cpu_threads=6)
+
+    assert cuda_backend.create_model.call_args.kwargs["cpu_threads"] == 6
+    assert cpu_backend.create_model.call_args.kwargs["cpu_threads"] == 6
+
+
 # === _transcribe_file() tests ===
 
 
@@ -527,6 +555,36 @@ def test_transcribe_file_openvino_gpu_midstream_fallback(mock_get_backend):
     assert tfr.model_path == "/mock/cpu/model"
 
 
+@patch("local_transcriber.transcriber.get_backend")
+def test_transcribe_file_midstream_fallback_preserves_cpu_threads(mock_get_backend):
+    ov_backend = _make_backend(
+        transcribe_error=RuntimeError("OpenVINO inference error"),
+    )
+    cpu_backend = _make_backend(
+        transcribe_result=_make_result(count=2, device_used="cpu"),
+        model_path="/mock/cpu/model",
+    )
+
+    def backend_for_device(device, **kwargs):
+        return ov_backend if device.startswith("openvino") else cpu_backend
+
+    mock_get_backend.side_effect = backend_for_device
+
+    with pytest.warns(UserWarning, match="Переключение на CPU"):
+        _transcribe_file(
+            model=MagicMock(),
+            actual_device="openvino-gpu",
+            backend=ov_backend,
+            model_path="/mock/ov/model",
+            file_path=Path("test.mp3"),
+            model_name="medium",
+            compute_type="fp16",
+            cpu_threads=6,
+        )
+
+    assert cpu_backend.create_model.call_args.kwargs["cpu_threads"] == 6
+
+
 @patch("local_transcriber.transcriber.get_backend")
 def test_openvino_gpu_strict_device_no_fallback(mock_get_backend):
     """strict_device=True + OpenVINO GPU ошибка → raise."""