D858 — Implementation PRs are reviewed by pr-reviewer before the coordinator reads them

Decision. The PRs that implement-issue and live-dev open go through the review-pr procedure (§3–§5): triage.sh gives the tier, standard and full tier get pr-reviewer rounds (at most three), and the coordinator verifies the findings and reads only the triage, the cited lines, the invariants ranges and supply-chain hunks. It no longer reads the whole gh pr diff; light tier is unchanged. Fixes are made by the author (the plan-implementer, or in live-dev the coordinator or a live-implementer), who only ever receives verified findings, and they are added as commits, never amended. The coordinator writes the PR's ## Self-review section after the last round.

Why. The full-diff read on the most expensive model is the cost D856 built review-pr to avoid, and the plan-implementer cannot spawn a reviewer, so its Self-review was never independent. live-dev used the client's generic /code-review, which does not know the area invariants and differs per client. The cost is one more round trip per PR and a coordinator that sees less code.

Alternatives rejected. - Keep the coordinator's full read and add the reviewer on top: pays for both reads. - Let the implementer spawn its own reviewer: not possible on every client, and an author's own reviewer is not an independent look. - Amend in fix rounds: rewrites the commit the next round diffs against.

Consequences. A fix round needs the implementer reachable (SendMessage, or a new plan-implementer with a fix mandate); the squash merge hides the extra commits. Supply-chain hunks are still read whole, by the user too (D857).