Problem

Sensei has advisory check 16 — cross-model description density — that validates:

1. Word count ≤60 (descriptions over 60 words dilute attention)
2. Lead sentence starts with unique action verb + domain
3. WHEN: with quoted phrases preferred over USE FOR:
4. Anti-trigger contamination warning (context-dependent)

Waza has none of these sub-checks.

Expected Behavior

Add a CrossModelDensityChecker to internal/checks/ that:

• Counts words in description, warns if >60
• Checks if first sentence leads with an action verb
• Prefers WHEN: format over USE FOR:
• Advisory only — does not change adherence level

References

• spboyer/sensei references/scoring.md — advisory check 16
• kvenkatrajan gist on cross-model pattern matching

Acceptance Criteria

• New checker in internal/checks/
• Word count check (>60 words = warning)
• Lead sentence heuristic
• Advisory status (not blocking)
• Tests