Merging findings Shipped
Findings merge on file and symbol rather than on who found them, so two specialists pointing at the same line corroborate instead of duplicating.
The merge key is the single authoritative definition of "two findings describe the same defect." It is keyed on file plus symbol, with a line-bucket fallback — and deliberately NOT on the specialist or specialty. Two lanes that independently spot the same problem collapse into one row that records both, which turns duplication into corroboration.
The incident that set the rules
A run on 2026-07-08 filed one blocking-loop issue five times and one path traversal six times across lanes. The duplication buried the unique signal. The fix was cross-cutting ownership: concurrency belongs to software-architecture, timeouts to networking-api, failure handling to reliability, missing tests to testing-qa, version hygiene to development-process. An issue outside your lane is still filed, but capped at low confidence as a cross-reference — unless the owning lane is absent from the run, in which case you are the coverage and file at full strength.
Review memory
A per-project database remembers what previous runs concluded. A finding you refuted does not come back at full confidence the next morning, and slices of the codebase that did not change carry their prior conclusions forward instead of being re-derived from scratch. Findings upsert on the run, the objective and the anchor — so re-running a review updates the record rather than starting a parallel one.