references/code-review.md

# Code review with reproducible findings

Read the diff and nearby code/tests; understand intended behavior and interfaces. Rank findings by impact: correctness, security/data loss, breaking changes, races, regression coverage, measured performance, maintainability.

Each finding includes: severity, exact grounded location, trigger conditions, impact, supporting code or test evidence and smallest fix. Separate demonstrated bugs from risks that require experiments. Avoid style-only blockers and fabricated line numbers.

A clean review means **no identified findings in inspected scope**, not proof that all software is correct. If review is limited by missing code or test execution, say so.

For data code examine grain/keys, cardinality, duplicate amplification, late events and replay. For AI code inspect tool side effects, authorization, structured output errors and external prompt content.