“Foundation models can hallucinate correct file paths and function names for SWE-bench problems without codebase context, indicating training data contamination.”