CodeBLEU Scored 91% for Both RAG Strategies While One of Them Hallucinated APIs 56.4% of the Time

arXiv 2609.12464·high signalXBlueskyLinkedInCopy link

For monolith-to-microservice translation, a Hierarchical Context-Resident Graph pipeline built with tree-sitter ASTs, a Spanner property graph, and a Gemini context cache dropped API hallucination from 56.4% to 16.2%, raised dependency resolution quality from 34.8% to 65.9%, and raised parent-child consistency from 26.7% to 45.5%. CodeBLEU rated both approaches at 91%, which is the point: text-overlap metrics masked structural failures behind syntactically plausible broken code. The graph approach is not free, since dense global context pushed the model into defensive over-engineering and cut cyclomatic complexity consistency from 71.6% to 46.7%.