Each repo is kept as a bare git mirror and parsed with tree-sitter (Python, JavaScript, TypeScript) into a graph of modules, classes, functions and methods with calls/imports/inherits/contains edges; every function is embedded into pgvector. A review diffs the PR against its merge-base, retrieves related code for each changed function by meaning, packs the diff and context into the model's 8k window with fixed per-part budgets, and asks for schema-constrained JSON findings. Validation then drops anything not on a changed line, not citing shown code, or not confident enough — with anchoring snapping findings that land 1–3 lines off back onto the change. The evaluation harness came before any improvement: 49 seeded bugs, 18 counterfactuals (the same change made safe by its callers) and 16 clean refactors, with an upgrade adopted only if F1's paired 95% bootstrap CI is above zero. Anchoring passed (F1 +0.113); four other ideas failed the gate and were not shipped — and the honest result that, with an 8B model, retrieved context doesn't pay for itself yet is reported as is.