Standard: sync (git fetch) before reviewing recent work; re-verify ground truth on user contradiction

Read standards/CLAUDE.md and compare its existing guidance with the proposed rule. Add language-agnostic guidance requiring git fetch and HEAD-versus-origin reconciliation, including reported SHAs and ahead/behind status, plus re-verification when contradictions recur. Done means all three acceptance criteria are satisfied in that file.

Goal

Add a rule to standards/CLAUDE.md that prevents reviews (and any "reason about recent work" task) from being performed against a stale or wrong-branch checkout, and that stops the agent from rationalizing contradictions instead of re-checking its inputs.

Background

In a live session, a full deep-review was run against a local main that was 15 commits behind origin/main. The agent never fetched. It then reviewed the tail two commits of a feature as if they were the whole feature, and produced confidently wrong conclusions:

• Claimed a feature was "flagged off / no UI wiring" — the wiring existed in unfetched commits.
• Filed two GitHub issues, one escalated to a "🔴 LIVE license violation" with a fabricated "config not in git" governance root cause — invented specifically to reconcile the user's pushback, rather than questioning whether the checkout was current.
• Only after the user asked "did you not pull from origin before the review?" did a git fetch reveal the 15-commit gap.

Both failure modes are generic (language/project-agnostic) and belong in team standards.

Proposed rule (for standards/CLAUDE.md)

Sync before reviewing; re-verify ground truth on contradiction.

• Before reviewing "recent work," a branch, a PR, or "the last N hours/days," always git fetch and compare local HEAD to origin/<base> before scoping the work. State the base and HEAD short-SHAs (and the ahead/behind count) up front. Never review a local checkout without first confirming it reflects the remote — for auto-deployed apps, origin/<default-branch> is what users are actually running.
• When a user contradicts your conclusion more than once, treat it as a signal that a foundational input is wrong (stale tree, wrong branch, wrong environment, wrong file) — stop and re-verify the inputs. Do not construct a new theory to explain away the contradiction on top of unverified ground truth.

Acceptance criteria

• standards/CLAUDE.md contains a rule requiring git fetch + HEAD-vs-origin reconciliation before any review/"recent work" task, with the base/HEAD SHAs reported.
• The rule includes the "re-verify ground truth when repeatedly contradicted; don't rationalize" guidance.
• Wording is project- and language-agnostic (applies in every repo a developer opens).

Notes

Complements the existing /deep-review skill (which already instructs fetching first but does not enforce it — see the companion skill-hardening issue).