Copilot code review — scope

This repository is reviewed by a separate policy reviewer — the .github/workflows/review-codex.yml workflow (the OpenAI Codex CLI on a ChatGPT subscription), which reviews every PR against the gamussa/coding-policy rules. It owns conventions and policy. Your job is the complementary lane: correctness and risk. Spend your review budget where the policy reviewer does not look.

Review for

• Logic errors, wrong conditions, off-by-one, incorrect edge-case handling.
• Race conditions, ordering bugs, non-idempotent retries, lost updates.
• Resource leaks (unclosed files/connections), unbounded growth, needless recompute on hot paths.
• Missing or wrong error handling that lets a real failure pass silently or crash — not stylistic preference.
• Security: injection, unsafe deserialization, authz gaps, secrets in code or logs, unvalidated untrusted input.
• Test coverage gaps: a changed branch or failure path with no test; an assertion that would pass even if the code were wrong.

Do NOT comment on (the policy reviewer owns these)

• Naming/style conventions, formatting, import order.
• Commit-message or PR-title format, changelog entries, branch naming.
• Rule/policy compliance, doc structure, skill/rule authoring conventions.
• Restating project rules — assume they are enforced by the policy reviewer.

Repo facts (avoid false positives)

• Verify a tool's real semantics before flagging — prefer a missed bug over a confident false positive.