chore(typecheck): clean datasets/eval/export/inference; enforce mypy on 12 packages

[xieofxie]

Summary

Extend the strict-mypy required-CI gate from 8 → 12 packages by cleaning datasets/, eval/, export/, and inference/ (189 errors → 0). Also restructures the dispatcher pattern in config/build.py to fix a regression timenick flagged on #825 (kwarg type-checking lost behind **dict[str, Any]).

What's in this PR

Dispatcher / overload cleanup (config/build.py, commands/config.py)

• generate_hf_build_config: added a third @overload accepting module: str | None and returning the union of the two narrow returns. Lets the generate_build_config dispatcher collapse to a single call (no kwarg duplication, no Any widening, no cast).
• commands.config command: ripped out the _shared_kwargs: dict[str, Any] = {...} workaround and replaced the if module: discriminator with isinstance(result, list) narrowing. Restores full call-site type checking against the callee's signature.

Cleaning 4 new folders (189 → 0)

Notable source-level fixes that cascaded:

• core/time_utils.format_timestamp_iso — added @overloads so float → str and None → None propagate; cleared 3 downstream [arg-type] chains.
• export/htp/base_writer.step decorator — replaced -> Any factory with Callable[[F], F], dropped no-op @wraps wrapper. One source fix cleared 18 cascading [untyped-decorator] errors across writers.
• export/htp/exporter — typed the None-init core fields with proper Optional annotations + single non-None assert in _apply_hierarchy_tags; cleared the union-attr chain.
• datasets/* — common pattern: class-level _image_col: str = "" default instead of runtime None-init in _detect_columns; eliminates Any | None property-return errors across image/segmentation/detection/depth datasets.
• eval/* — BaseImageProcessor | None guards across evaluators; pipeline() call-overload suppressions for runtime task strings that can't be statically matched against transformers' 60+ Literal overloads.

Configuration

• pyproject.toml ignore_missing_imports: added soundfile, sklearn.*, evaluate(.*).
• Dev deps: added scipy-stubs, types-psutil.
• .github/workflows/lint.yml: switched dependency install to uv sync --all-extras --all-groups so type stubs in [dependency-groups] dev are present in CI. Extended the required mypy loop from 8 to 12 packages.

Tracked follow-ups

• Remove dead reference to inference.onnx_config.* subpackage in HTP config generator #859 — two unguarded winml.modelkit.inference.onnx_config.* imports in export/htp/config_generator.py reference a subpackage that doesn't exist in the repo. Silenced with # type: ignore[import-not-found] referencing the issue; Remove dead reference to inference.onnx_config.* subpackage in HTP config generator #859 tracks delete-vs-restore decision.

🤖 Generated with Claude Code

[xieofxie]

revert of https://github.com/microsoft/winml-cli/pull/789/changes#diff-45b62ef5ac66eb35c9a46086706b3be0ecf8f6ab0f44c1eaad14df18858ab3a9

[xieofxie]

revert as comment in #825

[DingmaomaoBJTU]

Type-checking PR review

Overall: Clean, well-scoped PR. The approach of progressively moving packages into the required mypy gate is sound. Found one structural correctness issue, two warnings about suppressions masking real problems, and a few low-priority notes.

Bug - dead runtime guard in tensor_similarity_evaluator.py

mypy proves the isinstance(model, WinMLCompositeModel) branch is unreachable given the function's declared parameter type. The two # type: ignore[unreachable] suppressions silence the warning and make the guard invisible to callers. Either widen the parameter type to WinMLPreTrainedModel | WinMLCompositeModel so the check is live, or remove it and document that composite models are excluded at the call site.

Warning - cast admits subtype incompatibility (evaluate.py)

cast("str | dict[str, str | Path]", config.model_path) with a comment admitting it is "subtype-incompatible-but-runtime-compatible" suppresses a genuine type mismatch. The right fix is to widen WinMLEvaluationConfig.model_path's declared type to match what WinMLAutoModel.from_onnx actually accepts, rather than casting at each call site.

Warning - dead try-branches in config_generator.py

Neither inference.onnx_config.input_generator nor inference.onnx_config.patterns exists at runtime. Both try-blocks always fall through to their fallback paths. The # type: ignore[import-not-found] suppresses mypy but the code inside the try-blocks is dead. Consider removing the unreachable import bodies entirely (leaving only the fallback logic) until issue #859 resolves the module's fate.

Warning - duplicate if TYPE_CHECKING blocks (pipeline.py)

Two separate if TYPE_CHECKING: blocks appear at lines 26 and 30. They should be merged into one block. As-is it reads as if the second was added without noticing the first.

Info - class-level empty-string sentinel (datasets/image.py)

Replacing None with "" as the pre-detection sentinel is fine since all checks use if not self._image_col: consistently. Worth noting: a subclass that accidentally skips _detect_columns() will inherit the class-level "" and fail with KeyError: '' rather than a more debuggable AttributeError. Not a blocker.

Info - @step decorator TypeVar bound (base_writer.py)

F = TypeVar("F", bound="Callable[..., int]") is stricter than the old wrapper. Any @step-decorated handler that does not return int will now get a mypy error. Confirm all existing handlers annotate -> int; if any return None, they'll surface as new failures once dependent packages are gated.

[zhenchaoni]

Heads-up on the dict(processor(...)) / dict(tokenizer(...)) wrappers added in datasets/{depth_estimation,image,object_detection,text,image_segmentation}.py:

These shallow-copy the returned BatchEncoding into a plain dict, which drops BatchEncoding-specific methods like .to(device) and the .encodings property. Should be fine here since the result is immediately consumed by dataset.map (which plain-dictifies anyway), but it's the one change in this PR with non-trivial runtime surface area. A single smoke run through the datasets path would confirm nothing downstream was relying on the BatchEncoding type.

[xieofxie]

Type-checking PR review

Overall: Clean, well-scoped PR. The approach of progressively moving packages into the required mypy gate is sound. Found one structural correctness issue, two warnings about suppressions masking real problems, and a few low-priority notes.

Bug - dead runtime guard in tensor_similarity_evaluator.py

mypy proves the isinstance(model, WinMLCompositeModel) branch is unreachable given the function's declared parameter type. The two # type: ignore[unreachable] suppressions silence the warning and make the guard invisible to callers. Either widen the parameter type to WinMLPreTrainedModel | WinMLCompositeModel so the check is live, or remove it and document that composite models are excluded at the call site.

Warning - cast admits subtype incompatibility (evaluate.py)

cast("str | dict[str, str | Path]", config.model_path) with a comment admitting it is "subtype-incompatible-but-runtime-compatible" suppresses a genuine type mismatch. The right fix is to widen WinMLEvaluationConfig.model_path's declared type to match what WinMLAutoModel.from_onnx actually accepts, rather than casting at each call site.

Warning - dead try-branches in config_generator.py

Neither inference.onnx_config.input_generator nor inference.onnx_config.patterns exists at runtime. Both try-blocks always fall through to their fallback paths. The # type: ignore[import-not-found] suppresses mypy but the code inside the try-blocks is dead. Consider removing the unreachable import bodies entirely (leaving only the fallback logic) until issue #859 resolves the module's fate.

Warning - duplicate if TYPE_CHECKING blocks (pipeline.py)

Two separate if TYPE_CHECKING: blocks appear at lines 26 and 30. They should be merged into one block. As-is it reads as if the second was added without noticing the first.

Info - class-level empty-string sentinel (datasets/image.py)

Replacing None with "" as the pre-detection sentinel is fine since all checks use if not self._image_col: consistently. Worth noting: a subclass that accidentally skips _detect_columns() will inherit the class-level "" and fail with KeyError: '' rather than a more debuggable AttributeError. Not a blocker.

Info - @step decorator TypeVar bound (base_writer.py)

F = TypeVar("F", bound="Callable[..., int]") is stricter than the old wrapper. Any @step-decorated handler that does not return int will now get a mypy error. Confirm all existing handlers annotate -> int; if any return None, they'll surface as new failures once dependent packages are gated.

Fixed and one issue created for you #859