Finalize FastAPI refactor: fix 5 runtime bugs and add API test suite

[Copilot]

The aiohttp→FastAPI migration had several bugs preventing the API from running at all. This fixes them and adds 38 endpoint tests to validate the implementation.

Bug fixes

• Model never instantiated (model/v2/__init__.py): stevedore returns a class; was passing it unwrapped to ModelWrapper → every method call raised TypeError: missing 1 required positional argument: 'self'. Added ().

• Dead async code in warm() (model/v2/wrapper.py): method was async def and referenced self._loop/self._executor/self._workers that were never defined. Made synchronous, removed unused asyncio/functools imports.

• get_root() crashes on disabled docs (api/__init__.py, api/v2/__init__.py): APP.docs_url[1:] raised AttributeError when docs are disabled (url is None). Also get_root did await get_v2_version(request) on a plain def — made get_v2_version async.

• JSONResponse missing content= kwarg (api/v2/predict.py): JSONResponse(ret) passes the dict as a positional arg, not content.

• Form/File fields broken under Pydantic v2 + FastAPI Depends() (api/v2/utils.py): pydantic.create_model() in v2 doesn't expose Form/File FieldInfo defaults through __init__.__signature__, so FastAPI treated everything as query params. Replaced with a class factory that builds an explicit inspect.Signature with the correct FieldInfo defaults:

def __init__(self, **kwargs): ...
__init__.__signature__ = inspect.Signature([
    inspect.Parameter("field_name", ..., default=fastapi.Form(...)),
    ...
])

Test infrastructure

• conftest.py: session-scoped test_app and client fixtures; mocks stevedore loading with an in-process TestModel.
• test_v2_api.py: 38 tests across 7 classes covering root, v2 version, model list, model metadata, predict (required/optional params, file upload, validation errors), debug, and OpenAPI/Swagger/ReDoc endpoints.

Pre-existing test bug fixed

test_loading_error patched get_available_models (doesn't exist) instead of get_available_model_names, and didn't reset the MODEL singleton — the test only appeared to pass because load_model() short-circuited on the already-loaded singleton.

[sonarqubecloud]

Quality Gate failed

Failed conditions
C Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE