[Qiong Wu (qiowu) (DingmaomaoBJTU)]

Summary

Enabling WinMLCG ETW timing currently disables ORT graph optimization and forces one intra-op thread, so traced GPU timing runs use different session settings from ordinary perf. Preserve the existing ORT defaults and session-factory/monitor overrides instead.

The explicit CPU-fallback and CGC lowering controls, provider/device validation, measured-window attribution and event-loss checks remain in place. Documentation and regression tests cover default, custom and monitor-supplied optimization/thread settings.

Validation

• 57 focused session-option, CGEP tracing and GPU attribution/replay tests passed.
• Ruff check passed for src/ and tests/; changed Python files pass Ruff format checking.
• Mypy comparison with origin/main reports the same 24 existing environment/import diagnostics and zero new errors (missing third-party stubs and namespace/import issues in this environment).
• Hardware retest completed for 10 previously failing NVIDIA CGEP GPU cases using this commit, Runtime 2.8.183, original ONNX/saved inputs, and the same allowed CPU-fallback policy: 4 recovered with 200/200 valid GPU samples, zero event loss and separate accuracy PASS; 5 retain measured-window attribution failures (separate accuracy PASS); 1 retains a compiler failure. No case timed out. These results do not imply GPU-only execution or resolution of the remaining failures.