ONNX shape_inference wrong-shape drift via duplicate local functions

This repository contains a minimal .onnx proof of concept showing that a checker-approved model with duplicate local-function ids can produce a silent wrong output shape contract.

Files

shape_drift_scan.onnx: minimal PoC
generate_poc.py: regenerates the PoC
reproduce.py: reproduces the mismatch across checker, shape_inference, and ReferenceEvaluator

Reproduction

python generate_poc.py
python reproduce.py

Expected highlights

checker ok
shape_inference reports output shape [2, 2]
runtime produces output shape [2, 1]