fix(reviewer): surface author rationale to reviewer agents

[AbirAbbas]

Summary

Reviewer agents were producing findings that contradicted explicit design rationale stated in PR descriptions, because the rationale never reached them. Three structural causes:

1. Intake/anatomy truncated the description aggressively. Intake gate kept the first 500 chars, intake fallback 1000, anatomy 500. A 1100-char PR body with the rationale in the second half was literally cut off before any model saw it.
2. The reviewer agent never saw the raw description. It received pr_narrative, risk_surfaces, and intake_summary — all digested summaries. The author's voice was laundered through two summarization layers before the reviewer looked at code.
3. The anatomy prompt explicitly tells the model to discount the description ("what the CODE says, not what the PR description says"). That framing is correct for diff-vs-description divergence detection, but it leaves the reviewer with no channel for author intent.

Changes

• intake_phase (gate + fallback) and anatomy_phase: bump description truncation from 500/1000/500 → 4000 chars uniformly. Catches typical thoughtful PR bodies without unbounded prompt growth.
• review_dimension: new pr_description parameter, plus a dedicated "Author's Stated Intent" section near the top of the assembled prompt. The framing tells the reviewer not to defer to the description — but if a finding contradicts a design choice the author explicitly justified, the reviewer must rebut the stated rationale on its merits, not flag as if the rationale wasn't given. Findings on points the description is silent about are unaffected.
• ReviewOrchestrator._run_parallel_review: wire self.pr_data.description through to review_dimension.

Background

Investigated against a real review where the reviewer flagged "add idempotency check before POST" and "silent failure on 403" — both contradicting explicitly-stated rationale ("POST is additive on purpose", "fail-soft is intentional") sitting in the PR body the reviewer never saw. The same root cause produces a class of bad findings on any PR whose author bothers to explain why.

Test plan

• Existing tests pass — 25 pass, 1 pre-existing unrelated cost_tracker failure on main baseline (not introduced by this PR)
• Imports clean (from pr_af.orchestrator import ReviewOrchestrator etc.)
• Manual: re-run a review against a PR with substantive design rationale in the body and confirm findings either get dropped or get rebutted on the merits when they contradict that rationale

🤖 Generated with Claude Code