P1-INFRA-006: Static Analyzer — Commutative Op Matching

Summary

Implement commutative operator matching in the static analyzer's subgraph pattern matching, allowing binary operators with swapped inputs (e.g., Add(a, b) matching Add(b, a)) to match correctly against EP-provided skeleton patterns.

Context

Many neural network operators are commutative (Add, Mul, MatMul in some contexts). The static analyzer currently requires exact input order matching against EP subgraph patterns. This causes false negatives — patterns that should match fail because inputs are in a different order than the skeleton.

From plans/release/0315_release_plan/P1_CHECKLIST.md (P1-INFRA-006).

Current State

• Subgraph pattern matching uses strict input order
• Commutative ops (Add, Mul, etc.) with swapped inputs don't match EP patterns
• Causes false "not supported" reports in the static analyzer

Desired State

• Commutative ops matched with input-order-independent logic
• Pattern matching tries both input orderings for commutative binary ops
• No false negatives due to input order differences

Acceptance Criteria

• Commutative op list defined (Add, Mul, Max, Min, BiasGelu variants, etc.)
• Pattern matching tries both orderings for commutative binary ops
• Validated: known commutative patterns in QNN/OpenVINO skeleton match correctly
• No regressions in existing pattern matching tests
• All existing tests pass

Technical Notes

• CARDINAL RULE: no hardcoded model assumptions — commutative op list must be config/spec-driven
• Standard ONNX commutative ops: Add, Mul, Max, Min, And, Or, Xor, Equal
• Test with real QNN subgraph patterns that have commutative ops

Related Files

• plans/release/0315_release_plan/P1_CHECKLIST.md — P1-INFRA-006