fix(review): load review rubric via CLAUDE.md import

[Lotte-Verheyden]

The automated review bot loads CLAUDE.md as context but not AGENTS.md — confirmed empirically via canary probes (the bot reported AGENTS.md's content was absent from its context until imported through CLAUDE.md). This is why the ## Reviewing Pull Requests rubric never reached the reviewer, producing generic reviews that missed authoring-principle violations.

Changes

• New root CLAUDE.md that imports AGENTS.md via @AGENTS.md. No content duplication — AGENTS.md stays the single source of truth. This also fixes the authoring path: Claude Code loads CLAUDE.md, not AGENTS.md, so skill authors now get the authoring rules too. Import confirmed working (the review bot reported AGENTS.md in its context via this path).
• Judgment clause added to the review section: read each changed file as its runtime consumer, treat the checklist as a floor not a ceiling, and never dismiss "this line is meaningless to the reader" as a style nit.

AGENTS.md is otherwise unchanged and remains the canonical doc for authoring agents.

[github-actions]

@claude review

[Lotte-Verheyden]

@claude review

[claude]

Looks good — the debugging probe (BANANA line) I flagged in my previous review has been removed, and the diff now cleanly does what it says: adds a CLAUDE.md that imports AGENTS.md via @AGENTS.md, plus a judgment clause in the review section.

What was reviewed: confirmed AGENTS.md:51 no longer contains the temporary probe; checked CLAUDE.md's intro prose isn't meaningful duplication of the @AGENTS.md import (ruled out); checked the new judgment clause's audience framing against this PR's own CLAUDE.md import (ruled out); this is a docs-only change with no code, security-sensitive, or CODEOWNERS-covered paths touched.

Extended reasoning...

Overview

This PR adds a new root CLAUDE.md that imports AGENTS.md via @AGENTS.md, and adds a judgment clause to the ## Reviewing Pull Requests section of AGENTS.md instructing reviewers to read each changed file as its runtime consumer would. Both files are meta/authoring-instruction docs, not application code.

Security risks

None. No code paths, secrets, auth, or permissions are touched. .github/CODEOWNERS does not cover either file, so no ownership bypass concerns either.

Level of scrutiny

Low-to-moderate is appropriate: this is a docs-only change, but it governs how this very review process loads its own rubric, so I checked it carefully for self-inconsistency and for the debugging artifact flagged in my prior review. That artifact (a 'BANANA' probe instruction at AGENTS.md:51) has since been removed in a follow-up commit (chore(review): remove temporary import-verification probe), resolving the one blocking concern from before. The remaining diff is small, self-contained, and does not duplicate content between CLAUDE.md and AGENTS.md beyond a one-line pointer.

Other factors

Bug hunting found no confirmed bugs this run; three candidate issues (redundant intro prose, audience-framing inconsistency in the judgment clause, removed 'do not restate' guidance) were raised and refuted by verifiers. I independently agree these are non-issues — none affect the correctness or usability of the instructions for their intended readers.