How LLM Evaluation Produces Comparable Numbers: Inside the LFORLA Reverse Engineering Benchmark