feat: enrich intelligence MCP response with computed component stats

[SNooZyy2]

Summary

• Enriched hooks_intelligence MCP tool response with 15 previously-missing fields
• Loads actual SONA, MoE, Flash Attention, EWC, and LoRA components to compute real stats
• Adds performance section with benchmark metrics
• Companion to null-guard display fix (separate PR)

Problem

The hooks intelligence command crashes because the display code expects fields like learningTimeMs, routingAccuracy, searchSpeedup, cacheHitRate, and a performance section — but the MCP handler never includes them. The response only contained availability flags and notes.

Fix

Instead of simple boolean availability checks, the handler now:

1. Loads each intelligence component (SONA optimizer, MoE router, Flash Attention, etc.)
2. Calls getStats(), getLoadBalance(), getSpeedup() to extract real metrics
3. Computes derived values (success rate, load balance score, cache hit rate)
4. Includes all computed fields in the response alongside existing fields

Test plan

• Run claude-flow hooks intelligence --format json — verify all 15 new fields present
• Run claude-flow hooks intelligence — full dashboard renders without crash
• Verify SONA metrics: learningTimeMs, adaptationTimeMs, avgQuality
• Verify MoE metrics: expertsActive, routingAccuracy, loadBalance
• Verify HNSW metrics: searchSpeedup, memoryUsage, dimension
• Verify embeddings: cacheHitRate
• Verify performance section: flashAttention, memoryReduction, searchImprovement, tokenReduction, sweBenchScore

Generated with claude-flow