Add Review Lens: guided pull-request review node

[dovvnloading]

Summary

Adds Review Lens, a first-party guided code review node: paste a GitHub PR URL, fetch its unified diff, read it as a guided group-by-group walkthrough, and work severity-tiered findings (red/yellow/gray) with a deterministic weighted scorecard - plus follow-up Q&A answered over the stored diff. Read-only by default; nothing posts to GitHub.

What Changed

• New \code_review\ node kind end to end: \CodeReviewState\ (\�ackend/domain/node_states.py), graph methods (\�dd/set/store/fetch/start/complete/fail/dismiss/append_qa), 31 wire fields (full diff text deliberately off-wire, lazy-fetched), contracts payload + regenerated TS, save/load round trip.
• New \graphlink_plugins/review_lens/\ domain package: PR URL parsing, diff fetch over the shared \GitHubRestClient, and \ReviewLensAgent\ - a port of the retired single-file Code Review plugin's rubric (weights, verdict gates, fallback heuristics) extended to multi-file diffs with walkthrough grouping.
• New dispatch mixin (\�ackend/agent_dispatch/code_review.py) + 7 scene intents (\�ackend/api/intents_code_review.py); picker entry via the first-party hatch (\plugins/review_lens/, new Validation & Delivery category).
• New \CodeReviewNodeView.tsx\ (Setup/Walkthrough/Findings tabs, verdict banner, tier pills, dismiss/copy, Q&A) on the shared card chrome/tokens, wired through SceneCanvas/sceneStore/minimap/filter labels/announcements.
• Tests: 42 new backend tests, 14 new view tests, toFlowNodes/minimap/announcement cases; deliberate amendments to the pinned intent count (179->186), wire-key set, wire budget (5120->5324B), and bundle ceilings (+2.8%/+1.6%, no new dependency).
• Security fix (pre-existing failure, unrelated to the feature): bump the locked pypdf 6.15.0 -> 6.16.1 for CVE-2026-84309/84310/84311 flagged by CI's pip-audit (hashes verified against PyPI + local download).

Why

Code review - not generation - is the bottleneck on agent-authored diffs (large PRs get rubber-stamped instead of read). GraphLink had the retired single-file reviewer and Gitlink's change proposals, but no PR-diff review surface. This closes that gap following the repo's own first-party-kind conventions.

Validation

• App launches (E2E boot-smoke passes against the real backend + built SPA)
• Relevant workflow was exercised manually (all flows covered by unit/integration tests; live-PR fetch needs a GitHub token + network, verified by mocked-dispatch tests)
• \python -m compileall -q .\ passes (run from the repo root - matches CI)
• \python -m pytest -q\ passes (run from the repo root - matches CI: 3130 passed, 19 skipped; perf suite 8 passed)
• \cd web_ui && npm run check\ passes (matches CI: schema/typecheck/lint/vitest/build/bundle)
• 
  px playwright test\ passes (5/5, incl. plugin-picker node creation)
• \python -m ruff check .\ and \python -m mypy\ pass

UI Notes

• Screenshots included if the change is visual -- not attached; the card reuses the Gitlink card chrome/tokens 1:1 (tab strip, field rows, banners, stat rows) with new tier pills + verdict banner. Happy to add screenshots on request.

Additional Context

Follow-ups deliberately deferred: per-finding Post-as-PR-comment (needs inline-comment position mapping + approval UX), local \git diff\ as a diff source alongside the GitHub API. Dismissal is server-persisted and undoable; re-fetch supersedes the old review rather than merging.

[dovvnloading]

Code review - Review Lens (agent/review-lens-code-review)

Verdict: looks good to merge. No blocking issues. All four CI jobs are green on the latest push (Python checks incl. pip-audit on the bumped pypdf lock, Frontend checks, E2E 5/5, Build check). I reviewed the full diff (44 files, +5617/-26) against the repo's own conventions: first-party-kind ladder, SDK hatch discipline, undo-classification gate, wire-contract/codegen flow, and the Gitlink card patterns the UI mirrors.

What holds up well

• Correct layering throughout: pure domain package (\graphlink_plugins/review_lens/), dispatch mixin with the established deferred-import patch seams, thin intent wrappers, domain-only mutations. The placeholder-claim race discipline and cooperative-cancel shape are faithfully carried over from Gitlink.
• Undo posture is exactly right (URL set + dismissal = A with
  ecord_command; fetch/run/ask/cancel = B run-lifecycle), and the classification gate + count pin were updated, not bypassed.
• Wire discipline is principled: full diff text off-wire with a monotonic version key (the FIX-6 precedent), scores coerced to the string-valued-dict lane codegen admits, save/load round trip covered by test.
• UI mirrors the Gitlink card chrome faithfully (shell/menu/tabs/draft discipline/lazy-fetch guards/memo comparators), severity tiers render as red/yellow/gray with the precise severity kept in metadata, and clipboard/diff rendering follow the inert-text pipeline rules.
• Pinned-guard amendments (intent count, wire keys + fallbacks, wire budget, bundle ceilings, lock bump) are all commented with measured numbers - the ratchet discipline intact.

Findings (all non-blocking, suggested follow-ups)

1. (Medium, follow-up) \codeReviewFiles[].patch\ rides the wire with no reader. Each per-file patch is capped (6KB) but there can be up to 100 files, so a maxed-out node puts ~600KB of patch text on every scene snapshot - the same snapshot-cost class \codeReviewDiffText\ was kept off-wire to avoid. Nothing renders it (Walkthrough shows groups/paths; the diff viewer uses the lazy fetch). Only server-side reader is the fallback heuristic's _first_path_matching. Suggestion: project \patch\ out of the wire row (keep it in state + save file), or cap aggregate patch bytes at fetch time.
2. (Nit)
   eview_engine.py::_normalize_findings\ builds \known_paths\ (L395-398) but never reads it - dead code, remove or use it to drop findings pointing at unknown paths.
3. (Nit) \diff_fetch.py\ L142-146 (\if not files_truncated and page > 1: pass) is a no-op block that only restates the flag's default - delete it.
4. (Nit) \ est_review_lens_domain.py::test_walkthrough_caps_groups_and_paths\ asserts \len(groups) <= MAX_PR_FILES\ (100) next to the real <= 8\ bound - the first clause is vacuous, drop it.
5. (Note, accepted) Fetch fires \onSetPrUrl\ (undoable) then \onFetchDiff\ (run-lifecycle, not undoable), so undo after a fetch reverts the URL but keeps the fetched diff. Same shape as Gitlink's local-root + tree calls - consistent with precedent, just naming it.
6. (Note, accepted) \session_load\ restores lists uncapped (caps live in \complete_code_review_run). Save path always caps, so only a hand-edited save file could exceed them - matches how every other kind's restorer trusts the saver.

Security posture - read-only default, per-finding dismissal instead of bulk actions, no GitHub writes, token via the shared allowlisted client, prompt-injection rule stated in both agent prompts, secret material capped and flagged when truncated. The pypdf bump (pre-existing CVE finding, verified hashes) is appropriately scoped into this PR rather than left red.

No action required from me on merge strategy - squash-merge per repo convention.