Add review-publisher skill

[MaxDall]

What

Adds the review-publisher agent skill: it reviews a Fundus publisher PR (a new publisher, or an added/changed parser version) by crawling live articles and verifying the extracted ArticleBody mirrors the real article — no dropped paragraphs, no leaked boilerplate — then drafts a single GitHub review.

Built on the skills framework from its companion PR #940, but independent of it: the scripts run standalone (python skills/review-publisher/scripts/review.py …) and don't import the installer.

How it works

A crawl-once-then-sweep flow is a hard gate on every publisher under review:

• scripts/review.py — the driver (crawl/sweep/show/adjudicate/status/payload). Every candidate the sweep surfaces must be explicitly adjudicated, and status must report READY before any verdict.
• scripts/_sweep.py — diffs the extracted body against the live document to catch missing/leaked content.
• scripts/_store.py — persists crawl/adjudication state across invocations.
• scripts/sampler/ — layout sampling, so the crawl picks structurally diverse articles instead of near-duplicates.
• SKILL.md / PLAYBOOK.md — the agent-facing entry point and step-by-step procedure.

Library change

Adds BaseParser.body_selectors() (src/fundus/parser/base_parser.py) — a public accessor returning a version's summary / subheadline / paragraph selectors. This gives external tooling (the skill) a supported way to read them instead of reaching into the private _*_selector attributes.

Testing

tests/test_review_skill.py pins the sweep gate's failure modes — above all, the ways it could report a silent false "clean".