Add derek skill — KB article quality reviewer

[hilram7]

Summary

Adds the derek skill — a KB article quality reviewer invoked with /derek <kb-file>. Derek reviews KB articles against kb_style_guide.md and flags issues that Vale and Dale do not catch:

• Frontmatter validity — all required fields present and correctly formatted (tags includes kb, knowledge_article_id starts with kA, products uses underscores not hyphens, keywords has 8–12 items)
• Article type and structure — identifies article type from the title and checks for the correct H2 headings (Resolution: Symptom/Cause/Resolution; How-To Instructions: Overview/Instructions; How-To Q&A: Question/Answer)
• Title format — checks format against article type, flags product names in titles
• Product name usage — full name on first mention, short name after, no unapproved abbreviations (e.g. NDC, NPS)
• Keyword and description quality — flags generic keywords, vague descriptions, and obvious missing terms from the article body

Changes

• .claude/skills/derek/SKILL.md — new skill file; reads kb_style_guide.md directly as its rule reference
• .claude/skills/derek/rules/ — 18 YAML rule files deleted (superseded by SKILL.md)
• .claude/skills/derek/references/rule-schema.yml — deleted

Testing

Tested across 8 KB articles covering all 4 article types (Resolution Error, Resolution Symptom, How-To Instructions, How-To Q&A):

• Product ID format — consistently caught hyphens vs underscores by cross-referencing the product names table in kb_style_guide.md
• Unapproved abbreviations — NDC, NPS flagged every time as non-approved short names
• Keyword quality — reads article body and suggests specific missing terms
• Structure + title mismatch — when title and structure conflict, derek explains both options for resolving it and notes that fixing the title changes the required headings
• Placeholder knowledge_article_id — correctly caught kA0Qk000000XXXXKAA as not a real ID
• Vale/Dale separation — articles with contractions were not flagged by derek
• Output format — table output confirmed across multiple runs; context reuse fix verified (two consecutive runs in same session produce identical fresh output, not "my analysis is unchanged")
• Structure consolidation — multiple missing headings for the same article type correctly consolidated into a single row with full context

Known Limitations

• Non-determinism — AI review produces slightly different results between runs; borderline findings (especially keywords-quality) may appear or disappear. Re-run if a finding seems unexpected.
• Run in a fresh session — running /derek multiple times in the same Claude Code session may cause context reuse. Open a new session for each review.