[DiadkoShmek]

Removing an Expand from PRelu's data input can turn a valid model into one rejected by full ONNX checking and strict shape inference. PRelu broadcasts the slope to the data shape; its other input cannot supply bidirectional broadcasting to justify that rewrite.

For example, x: float[3] expanded to [2,3] with a float[2,3] slope is valid. The current rewrite removes Expand, after which shape inference reports rank 1 against the declared rank 2 output. Omit only the first/data-input PRelu rule, retaining slope-side removal and every other operator's registrations.

Six regressions fail full validation on the original rule, and the safe slope-side control passes. After the change all 21 tests in the rule module pass. An independently generated 124-case check covers rank/singleton growth, symbolic and unknown dimensions, runtime targets, shared graph outputs and Add/Mul controls; all fixed graphs pass both validators. lintrunner -a and git diff --check pass.

This concerns the exported opt-in expand_before_binary_op_rules, not the default optimizer rule list. ONNX Runtime permissively computes the same values for the invalid original rewrite; no numerical runtime failure is claimed. Conservatively retaining a redundant data-side Expand is intentional.

AI assistance: OpenAI Codex investigated, implemented and independently checked this change with pinned local Python/ONNX/ORT dependencies. No human code review, full-suite validation or upstream acceptance is claimed.