Read the architecture and trace changes across service
adapters, packages/kumo-core, PQL, and DiskGraph.
- Preserve time boundaries and prevent leakage. Check empty-input behavior and numerical results when changing aggregations, sampling, or training.
- Check consumers of changed output schemas, model artifacts, module exports, defaults, and feature flags. Consider persisted data and cached statistics.
- Reuse existing helpers and keep validation with its owning component.
- Add focused regression coverage for behavior changes; explain removed tests and restrictions. Keep unrelated changes separate.
Use ML review patterns for concrete examples, query orchestration guidance for service rules, and PQL instructions for parser changes. Follow the DiskGraph source boundary for native-engine changes.