[mgratzer]

The default review now asks where else a constant, bound, vocabulary or rule is written and what holds the copies together, so a fact restated in another tier or language is found at the diff, not after it lands.

Notes

• Judgment calls (HITL Issue, run unattended): the question uses the Issue's proposed wording verbatim, as a fifth Design question. The cut is the Issue's first candidate: item 5 drops "hardcoded values that should be constants", which the new question covers from the other side (a value worth a constant is one written in more than one place). Net words still grow by about 30 across the two modules; folding it into the sibling question would have made that question carry two searches.
• The rubric's P1 examples gain "a constant, bound or rule restated elsewhere with nothing holding the copies together", beside the copied module.

Exercised

• forge-new:forge-reflect (this branch) on a sandbox branch adding GET /notes/:id/preview plus a client that restates the 80-char length (with a "keep in sync" comment) and the id regex (without one) → P1 on the restated length, naming the id regex at src/server.js:17 and the third copy in docs/api.md; plus 4 P2s
• forge-old:forge-reflect (base 24f900e), same branch → P1 on the restated length only, no mention of the id regex; same 4 P2s, plus a tracker-spec P1 (the new run raised it as an open question) and one on the fixture's commit title
• Not exercised: a full forge-ship --guard run; the sandbox branch stayed local, no sandbox PR opened

Plan

Plan as first made (no approver: run unattended, started by another session with --guard)

• Design block: interface is one Default Pass question in _shared/review-dimensions.md item 3; it hides nothing new. The same decision already lives in the sibling question (feat(skills): ask where a decision already lives, in the plan, the review and the risk verdict #154) and the rubric's "a module or sequence copied from one that exists".
• Phase 1: add the question after the sibling question, drop item 5's "hardcoded values that should be constants", add the rubric example, run scripts/sync-shared.sh and scripts/check-skills.sh.
• Phase 2: exercise with forge-reflect from branch and base snapshots on a sandbox branch that restates a fact.
• Scope: not _shared/risk-rubric.md (feat(land): a gate loosened so the diff passes needs review, like a weakened test #157 edits it); no new module or variant.

Deviations (as of c0b492e): none

Closes #156

Risk

Verdict: needs review, judged on c0b492e

• Peer review: none configured
• Size: 55 counted lines (0 excluded) → ok
• Always-review paths: none
• Tests: ok
• Exercise: ok
• Plan: ok
• Design: ok
• HITL: feat(review): ask where else a fact is written and what holds the copies together #156 shipped without an approved Plan, judgment calls listed
• Trust: ok
• Deployment: none
• What the checks would miss: The cut the Issue left to the maintainer. Item 5 loses "hardcoded values that should be constants", which also covered a single magic number written only once, and the new question does not cover that case. The new P1 example could also raise how often reviews block on hand-kept mirrors that are intentional.

Where to look (needs review only)

• skills/_shared/review-dimensions.md:9: The Issue left two cuts to choose from. Is adding a fifth Design question (about 30 more words) better than folding this into the sibling question, as the PR's Judgment calls Notes bullet argues?
• skills/_shared/review-dimensions.md:16: With "hardcoded values that should be constants" gone from item 5, will a review still flag an unexplained magic number that appears only once?
• skills/_shared/review-rubric.md:8: Should a restated fact with nothing holding the copies together be a P1 ("should fix before merge"), or does that over-block mirrors that are only kept in sync by a comment?