review

Target: $ARGUMENTS (`diff` = uncommitted + unpushed work, the default; `repo` = the whole codebase; a path = that subtree). `--no-fix` stops after round one with findings only. ## Round one Establish the target and state it — `git status --porcelain`, `git log --oneline origin/main..HEAD`, `git diff --stat` — so the reviewers are not guessing what changed. Run in parallel, each on the same target, each in its own fresh context: - **go-reviewer** — correctness and the norms `.golangci.yml` deliberately does not lint. - **honesty-auditor** — every figure a reader could act on: layer, provenance, confidence, denominator scope, error bars, the refusals. - **surface-auditor** — whether the published prose still describes this binary. - **store-steward** — only if a migration, a stored field or a write path is in the target. **Edit nothing while they run.** A mid-review edit changes the tree under the other reviewers. Merge the findings: drop duplicates and anything the linter already catches; rank by whether a user would act on a wrong result. For each survivor: file:line, the concrete failure (input → wrong output), the smallest fix. Say which findings you could not verify. ## Rounds two and three (unless `--no-fix`) Fix the confirmed findings, run `bash .claude/checks/gate.sh`, then send **only the fixed hunks** (`git diff` of those files) back to the reviewer that raised each finding, asking whether it is closed. A finding still open after round three goes into the wo...