feat(pr-review): extract review rubric to files, 4-lens restructure

[Troublor]

Summary

Refactors the shared claude-pr-review action so the review rubric is maintainable and the review agent stays at a structural altitude.

• Extract the rubric to a sibling rubric.md, injected into the prompt at compose time (cat into GITHUB_OUTPUT, same mechanism as the existing orchestration var). The rubric is now edited and reviewed as prose instead of a YAML block scalar with heredoc-avoidance hacks. No change to what the agent receives at the always-loaded layer.
• Restructure the 8 priority items + the 6-item "also check" list into 4 leverage-ordered lenses — design & data structures, correctness & safety, surface stewardship, change & operate — so the reviewer reasons by mode instead of keyword-matching a flat checklist (the documented cause of nit avalanches). Every prior check is preserved, just folded under a higher-altitude header.
• Add rubric-detail.md for conditional, line-level checks (persistence/crash-consistency, unit-safety, test & perf surface). It's staged into $GITHUB_WORKSPACE/.claude-review/ and the rubric routes the agent to read it on demand, only for the areas a given diff touches — true progressive disclosure, kept out of the always-loaded prompt.

The posting-contract phases and severity scale stay inline (they carry ${{ github.* }} interpolations / are posting mechanics, not rubric content). Scope is limited to this action; the README's stale mega-agents action paths are a separate bug to fix in a follow-up.

Test plan

• This PR is self-dogfooding: documentation's own claude.yml runs pr-review via the local action path, so the modified action reviews this very PR. A green review run exercises the compose injection (rubric.md), the workspace staging (rubric-detail.md), and the new 4-lens prompt end to end.
• action.yml validated locally: composite run-script passes bash -n; block-scalar indentation and the rubric/orchestration multiline GITHUB_OUTPUT emits are consistent.
• Confirmed no review check was dropped relative to the previous inline rubric (incl. unsafe/SAFETY, concurrency hazards, public-API rationale, framework-convention respect).

---

This PR was generated by an automated agent.

[claude]

Clean.

• Reviewed the claude-pr-review action refactor: rubric extraction into rubric.md, 4-lens restructure, and on-demand rubric-detail.md staged into the workspace .claude-review/ directory.
• Verified all prior rubric items (data structures, design, correctness, complexity, breakage, tests, naming, style; plus unsafe/SAFETY, errors, concurrency, public-API rationale, framework conventions, logging, doc proximity, duplication) map into the new lenses or Do NOT flag - nothing dropped.
• Compose mechanics mirror the existing orchestration pattern (set -euo pipefail, randomized heredoc delim, action-path source); action inputs are unchanged, so callers are unaffected.
• Title follows Conventional Commits; description explains what and why with a self-dogfood test plan.
• No new actionable findings.
• Still open from earlier reviews: none.

[claude]

Re-review update

• Reviewed incremental changes since prior review (commit c43023a): prettier formatting on rubric.md and rubric-detail.md — added blank lines between headers and content, plus italics asterisk-to-underscore conversion.
• Per the rubric, formatting owned by the repo formatter is not a review concern.
• No new actionable findings. Re-review verdict matches prior round: clean.
• Still open from earlier reviews: none.