[Michael-Duck]

Summary

The package used functions where a type was expected: callable as an annotation in Extractiveness.__init__ and in find_executable_batch_size, np.array (the constructor function, not the array type) as the return annotation of DropMetrics._align_bags, and a bare Annotated as the type of Arg.type. Type checkers reject all of them and typing.get_type_hints() cannot resolve them at runtime. This replaces them with typing.Callable, numpy.ndarray and typing.Any, and adds a regression test that scans every module of the package for annotations that do not refer to a type.

Notes

Only annotations change, runtime behaviour is unchanged. The new test is a static (ast-based) check: it needs no model, no GPU and no network access. Scope: this fixes the class of errors the issue points at - annotations that are not types - and guards against new ones; the ~1.2k remaining mypy errors reported in the issue are unrelated to it.

Tests

• python3 -m pytest tests/unit/test_type_hints.py -q -> 285 passed. On the unmodified tree the same test reports 4 failed, 281 passed - one failure per file changed above: cli_args.py (bare Annotated), metrics/harness_compatibility/drop.py (np.array), metrics/metrics_sample.py (callable), utils/parallelism.py (callable).
• ruff check src tests and ruff format --check src tests -> clean. The whole-repo ruff check . and ruff format --check . (what make quality runs) are clean too.
• Wider suite: python -m pytest tests/ -q --ignore=tests/slow_tests -> 953 passed, 17 failed, 5 skipped. Those 17 failures are pre-existing and unrelated to this change: the identical set fails on the unmodified origin/main tree. They are 8 failures in tests/unit/tasks/test_registry.py (TypeError: HfApi.list_models() got an unexpected keyword argument 'model_name', a huggingface_hub API change) and 9 in tests/unit/models/vllm/test_vllm_model.py + tests/unit/utils/test_caching.py (vllm is not installed in the sandbox).
• tests/slow_tests was not run: it downloads models and needs a GPU.

Fixes #759