fix(reviewer): surface author rationale to reviewer agents

[AbirAbbas]

Summary

Supersedes #25 — same fix, re-applied onto current main. The old branch predated a history rewrite, so GitHub showed a phantom ±40k-line diff; the actual change is this single commit (+33/−3), cherry-picked from 935e4c1 with one trivial conflict resolved.

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

1. Intake/anatomy truncated the description aggressively (500/1000/500 chars) — rationale past the cutoff was chopped before any model saw it.
2. The reviewer agent never saw the raw description — only digested summaries (pr_narrative, risk_surfaces, intake_summary).
3. The anatomy prompt explicitly discounts the description ("what the CODE says, not what the PR description says") — correct for divergence detection, but it left reviewers with no channel for author intent.

Changes

• intake_phase (gate + fallback) and anatomy_phase: bump description truncation 500/1000/500 → 4000 chars uniformly.
• review_dimension: new pr_description parameter with a dedicated "Author's Stated Intent" section near the top of the prompt. Framing tells the reviewer not to defer to the description — but a finding that contradicts an explicitly-justified design choice must rebut the stated rationale on its merits. Findings on points the description is silent about are unaffected.
• ReviewOrchestrator._run_parallel_review: wire self.pr_data.description through to review_dimension.

Conflict resolution vs. old #25: main has since added feedback_section (human HITL guidance) at the same prompt-assembly insertion point — both sections are kept, human guidance first.

Test plan

• Prompt-capture harness against a mocked router.app: with a 5000-char description carrying a rationale marker at char ~2400, the marker now reaches the intake gate, intake fallback, anatomy, and reviewer prompts; the reviewer prompt contains the "Author's Stated Intent" section capped at 4000 chars; empty description produces no section; human-guidance section still renders and precedes it. Same harness on main fails exactly these checks (10/10 pass on this branch, 4 relevant failures on main).
• ruff check src/ scripts/ clean
• pytest tests/ — 73 passed
• go build ./... && go vet ./... && go test ./... + gofmt clean
• docker build succeeds

---

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[AbirAbbas]

Follow-up pass to make this merge-ready (merged current main + four commits):

• Collision-safe description embedding: the author-controlled description was interpolated inside a fixed ``` fence — a description containing a fence could close the block and leak text into the instruction channel. All four embedding sites (intake gate, intake fallback, anatomy, reviewer section) now wrap it in collision-proof tags computed against the content, with explicit treat-as-data framing.
• Go node parity (the default install since Installing this repo gives you the Go node, as pr-af #64): description caps 500/1000 → 4000 in the Go intake/anatomy prompts, and the "Author's Stated Intent" section — same wording, same placement after the human-guidance section — threaded through ReviewDimensionInput → runParallelReview, so spawned and coverage-gap dimensions get it too.
• Committed tests on both sides: the PR-body capture harness is now tests/test_description_prompt_contracts.py plus Go prompt-contract tests and regenerated goldens: marker-at-char-2400 reachability, 4000-char cap, empty description omits the section, ordering after the feedback section, and fence-collision safety.

Gates: ruff clean, pytest 78 passed, go build/vet/test/gofmt clean.