intent-reviewlisted

# Intent Review Run **one** review that compares a finished diff against the **original request, verbatim**, and route every surviving finding to a fix, an escalation, or a durable record. This is the **third oracle** in the review chain — independent of, and blind to, the plan. **Why this stage exists.** The chain already has two oracles, and both can pass while the result is still wrong: - **spec review** asks *"does it match PLAN.md?"* — but the plan itself can mis-encode the intent. - **correctness review** asks *"does it run correctly?"* (blind to the plan) — but correct code can still build the wrong thing. Neither asks *"would the person who asked for this recognize it as what they asked for?"* When intake or design misread the intent, every downstream gate passes **consistently** and the result is still wrong (Goodhart: the `<verify>` commands are written by the plan author, so they only measure what the plan thought mattered). The single human gate after implementation is the PR merge — but nothing gives the merger an intent artifact to check against. This stage is that check, run by a fresh blind reviewer before hand-off. **Three oracles, mutually blind:** | Oracle | Skill | Asks | Blind to | |---|---|---|---| | PLAN | spec review (per-task) | does it match the spec? | — | | runtime | `/correctness-review` | does it run correctly? | the plan | | intent | `/intent-review` (this) | is it what the user asked for? | the plan | `correctness-review` is blind to t