Health Check

Knowledge Base vs Code:

• For each concept article mentioning a feature: grep for the feature name in prepare.py and train.py. Does the implementation match the article?
• For each decision: is it reflected in current code?

Staleness:

• Are any knowledge articles' last_updated dates more than 7 days old while their source files have been modified more recently? (Check with git log)
• Are any CLAUDE.md gotchas no longer relevant?

Internal Consistency:

• Do cross-links in knowledge articles point to articles that exist?
• Are there duplicate articles covering the same topic?
• Do decision statuses make sense? (e.g., "accepted" decision that was later superseded should be "superseded")

3. Suggest improvements

For each FAIL:

• What is inconsistent
• Which source is likely correct (prefer: code > spec > knowledge > CLAUDE.md)
• Suggested fix

Also suggest:

• New concept articles for topics discussed in council reviews but not yet in the knowledge base
• New experiment articles for completed experiments not yet summarized
• Interesting connections between existing articles

4. Output

Write the report to knowledge/HEALTH_CHECK.md:

# Knowledge Base Health Check

**Date:** <today>
**Status:** X pass / Y fail / Z suggestions

## Failures
### F1: <description>
...

## Suggestions
### S1: <description>
...

Return a 1-2 sentence summary to the caller.