“Large language models can answer questions about the VS Code codebase reasonably well without explicit context because they have learned its structure directly into their weights from seeing the public code during pre-training.”