feat(resolution): semantic merchant clustering + correction learning

[kayodebristol]

Summary

Phase 3 intelligence layer for the resolution engine.

Requirements

• Embed merchant/entity descriptions (via ai-providers or local model)
• Cluster semantically similar transactions (not just text matching)
• Learn from user corrections: when user re-categorizes, update resolution model
• Generate explanation objects: "Classified as Groceries because: similar to previous Trader Joe's transactions, amount range matches, weekly pattern"

Acceptance

• Semantic clustering finds related merchants that text matching misses
• User corrections improve future resolution accuracy
• Explanations are generated for all resolutions

[kayodebristol]

@copilot Please implement this issue.