Review Rubric

<!--
review-rubric.md — Chapter 8. The senior-review prompt that turns a cheap
executor's diff into a graded verdict.md, then a ship / send-back call. ONE pass,
no loops: Fable reviews once, you read the verdict, you decide.

Run it against the diff:
    git diff main > change.diff
    claude --model fable --effort high "$(cat prompts/review-rubric.md)
    $(cat change.diff)"

Then collapse the verdict to a decision with the shell gate:
    bash tools/gate.sh verdict.md
A worked verdict ships at examples/verdict.sample.md (1 blocker → SEND BACK) and
examples/verdict.clean.md (no blocker/major → SHIP).

Route the review the same way you route the work — high-stakes paths only:
    bash tools/needs_senior.sh examples/changed.txt
-->

Fresh context: you did not write this diff and you do not defend it.
Try to prove it does NOT meet the spec; a pass you can't break is the only pass.

Review this diff as a senior engineer. Do NOT edit anything; produce a verdict
only. Grade every finding:
  BLOCKER (ships broken / security / data loss / breaks done-criteria)
  MAJOR   (real bug under a plausible input, tests missed it)
  MINOR   (missing edge case / weak error handling)
  NIT     (style / naming, no behavior impact)
For each: severity, file:line, one-sentence why, the concrete fix.

Check the diff against the handoff's constraints and gate — did the executor
edit only the allowed files and honor every frozen constraint? And check the
GOAL, not just the code: does this diff accomplish the plan's actual goal, or
just pass its gate? Clean code that solves the wrong problem is a finding.

End with VERDICT: SHIP or SEND BACK (list exactly what must change).

<!--
The gate the verdict collapses to: any Blocker or Major → SEND BACK, else SHIP.
Nits and Minors never block. The verdict is an input to your decision, not the
decision itself — go to the file:line and confirm a Blocker is real before you
route a fix. A review that's always right is a review you've stopped reading.
-->