BRIDGE-1: Ship ← Digital Twin Cache (10.5M pre-classified products)

Modify the Ship service to check Supabase product_digital_twins (10.5M records) BEFORE calling LLM for classification. This is the single highest-impact optimization. Flow: 1. Embed incoming product description (text-embedding-005, 768-dim) 2. RPC call: match_digital_twin(query_embedding, threshold=0.15, limit=1) 3. If match with confidence >= 0.85: return cached classification (cost: $0.001) 4. If no match: fall through to Oracle KB (Bridge 2), then LLM (T2) 5. After LLM classification: WRITE BACK to product_digital_twins (grows cache) Supabase RPC to create: ```sql CREATE FUNCTION match_digital_twin(query_embedding vector(768), threshold float, match_limit int) RETURNS TABLE (pattern_id text, hts_code text, confidence float, ...) ``` The product_digital_twins table already has 10.5M rows with embeddings. The datalake_flow.py pipeline continuously adds new patterns. Revenue impact: 80-90% cache hit rate. Cost per classify drops from $0.03 to $0.003. Output: Modified ship/main.py with twin cache lookup + Supabase RPC migration Blocked by: Ship service (06c034a5) must exist first