type: clean mypy in optracing, quant, serve, session, sysinfo, telemetry, utils and enforce in CI

[xieofxie]

@

Summary

Brings seven more winml.modelkit packages to zero mypy errors under the strict [tool.mypy] config and wires them into the required CI type-check gate (16 → 23 packages). Continues the incremental cleanup from #896.

Cleaned: optracing, quant, serve, session, sysinfo, telemetry, utils (sysinfo was already clean). Fixes follow the repos typing playbook — fix at the source first; cast() only at genuine untyped third-party boundaries; Any only for true dynamic pass-throughs.

Real bugs surfaced (not just typing)

• utils/hub_utils.py version import — from ..version import __version__ referenced a non-existent module; guarded by try/except, so the export version was silently always "unknown". Fixed to from .. import __version__. (The containing helper inject_hub_metadata is dead code — tracked in Clean up unused inject_hub_metadata in utils/hub_utils.py #913.)
• utils/hub_utils.py — model_info.modelId → model_info.id (huggingface_hub renamed the attribute).

Notable fixes

• session — bound a non-None local after auto-compile so narrowing survives into the lambda/comprehension; explicit guards for resolved_ep is None and single-mode model_path; widened PDH counter containers to float; class-level _initialized annotation; distinct ctypes struct locals.
• telemetry — annotated the LoggerProvider/Logger instance attributes (were inferred as None, cascading into unreachable/attr-defined); OTel/JSON boundary casts; typed the @contextmanager/@asynccontextmanager generators.
• serve — cast() at json.loads/app.state boundaries; import binascii to use binascii.Error (typeshed doesnt expose base64.binascii); explicit single-mode model_path guard.
• utils — removed stale # type: ignores; DataclassInstance cast for generic-TypeVar dataclass reflection; EPAlias key cast for the alias lookup; loop-variable rename to fix InputTensorSpec/OutputTensorSpec shadowing.

Config changes (pyproject.toml)

• qairt.* → ignore_missing_imports (external Qualcomm AI Runtime SDK; imported only inside compile_qairt_bin.py, which runs in a separate venv-winml subprocess — absent from the CI env).
• windowsml → follow_untyped_imports = true (installed but ships no py.typed; analyzing its source is clean and keeps its inline annotations instead of collapsing to Any).

Verification

• Combined required-gate invocation (all 23 packages): Success: no issues found in 300 source files.
• ruff clean; unit tests green for the touched packages (2 pre-existing OpenVINO-EP-unavailable failures in this env, confirmed unrelated).
• Branch-vs-main diff is content-only (line endings normalized to LF to match main).
  @

[DingmaomaoBJTU]

All five findings from the previous round are addressed in f2091a2:

• Bug fixed: �xporter.py timestamp now uses explicit is not None guard — epoch=0 is handled correctly.
• Warning fixed: data_utils.py signature widened to Mapping[str, Sequence[int | str | None]]; # type: ignore[unreachable] removed.
• Warning fixed: _pdh.py running-time counters reverted to dict[str, int] with explicit int() at assignment; delta computation is now clean.
• Info addressed: session.py InferenceError context now includes onnx_path and device.
• Info addressed: cli_api.py cast no longer includes None in the union type.

LGTM.