[Asuka (akasakariko)]

Description

• Treat inserted FP32-to-FP16 Casts as narrowing when checking whether a model-authored consumer Cast can bypass them
• Keep the precision-loss exemption for inserted consumer Casts so adjacent CPU FP32 fallback operators can still share FP32 intermediates
• Add a numerical regression that forces the MatMul fallback path even on CPUs with native FP16 support and checks both FP32 and FP16 outputs with optimizations disabled and extended optimizations enabled

Motivation and Context

Fixes #33998

The duplicate-Cast remover treated inserted downcasts as precision-free, allowing an explicit FP16-to-FP32 Cast to read the unrounded FP32 fallback result

For FP16 inputs representing 0.7, the FP32 output was 0.49027347564697266 instead of the FP16-rounded 0.490234375

Validation

• Reproduced the issue attachment on Linux x86_64 with ONNX Runtime 1.31.1 under both ORT_DISABLE_ALL and ORT_ENABLE_EXTENDED
• Built onnxruntime_test_all in Release mode on macOS ARM64
• Confirmed TransformerTest.Fp16MatMulFallbackPreservesRoundingBeforeCast fails on the unchanged implementation at both optimization levels and passes with this fix
• ./onnxruntime_test_all --gtest_filter='TransformerTest.*' — all 31 tests passed
• lintrunner onnxruntime/core/optimizer/insert_cast_transformer.cc onnxruntime/test/framework/insert_cast_transformer_test.cc — passed
• git diff --check — passed
• Full test suite and Linux x86_64 source build not run locally

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.