feat: split writing reviewer into zh and en jobs

[lipingtababa]

Summary

• Split review-writing into two separate jobs: review-writing-zh and review-writing-en
• review-writing-zh: reviews only zh/ files using Chinese-specific AI fingerprint rules, outputs a Chinese report
• review-writing-en: reviews only en/ files using English-specific AI fingerprint rules, outputs an English report
• Each job skips gracefully if no files in its language are changed in the PR
• review-facts remains unchanged, covers both languages

Test plan

• Open a PR with only zh/ changes — only review-writing-zh produces a substantive comment
• Open a PR with only en/ changes — only review-writing-en produces a substantive comment
• Open a PR with both — both reviewers fire

[lipingtababa]

Writing Reviewer Split: Informational Review

Good structural improvement — splitting the monolingual reviewer into language-specific jobs is the right call. A few observations:

What works well

• Graceful skip logic: Each job checks whether its language directory has changes and exits early if not. Clean approach.
• Language-appropriate output: The zh job produces a Chinese report with Chinese review criteria; the en job stays in English. This makes review comments actionable for contributors working in each language.
• Fact-checker refinement: Adding file to the output format template for review-facts is a small but useful improvement now that it explicitly covers both directories.

Observations

1. Duplicated prompt structure: The zh and en prompts share identical review dimensions (AI fingerprint detection, terminology, logic, information density, structure, style) but are fully duplicated rather than templated. This is ~120 lines of parallel content that must be kept in sync manually. If the review criteria evolve (e.g., adding a new dimension), both prompts need updating. Worth considering whether a shared base prompt with language-specific overrides would be more maintainable — though the simplicity of full duplication has its own value for a file this size.

2. Sticky comment collision risk: Both review-writing-zh and review-writing-en use use_sticky_comment: true. If claude-code-action keys sticky comments per workflow (not per job), one job's comment could overwrite the other's. Worth verifying the sticky comment keying behavior — if it collides, setting a distinct sticky_comment_id per job would fix it.

3. Chinese AI fingerprint list is thorough: The zh prompt includes a well-curated list of Chinese AI writing fingerprints (corporate buzzwords like 赋能/闭环/抓手, structural patterns like symmetric parallel lists). This is more specific and useful than a generic "check for AI-sounding text" instruction.

4. Minor: The en prompt still contains the line "Space between CJK and Latin characters" (line ~101 area). Since this job only reviews en/ files, that rule is unlikely to trigger, but it's slightly misleading in an English-only context.

Verdict

Solid, focused change. The duplication and sticky-comment-key questions are the main things to keep an eye on.