#333 Sherlock Phase 3: Implement Semantic Idempotency

PHASE 3: Semantic Idempotency (Efficiency) IMPLEMENTATION: Option 1 - Embedding-based Semantic Similarity ARCHITECTURE: 1. Create tools.semantic.embed - Generate embeddings via Ollama (nomic-embed-text) 2. Create tools.semantic.compare - Cosine similarity comparison 3. Add semantic_check parameter to ctx.step() - opt-in only 4. Store embeddings in checkpoint state alongside result 5. On replay: compute new embedding, compare similarity, log result 6. MUST run in Activity Workers only (avoid DB connection exhaustion) STORAGE: - Checkpoint state: {result, schema_hash, embedding, similarity_threshold} - Comparison logged: {old_embedding_hash, new_embedding_hash, similarity_score, passed} WORKFLOW EXAMPLE: builder.activity( task_id='summarize_text', function='tools.llm.call', kwargs={...}, semantic_check=True, similarity_threshold=0.95 ) REQUIREMENTS: - All verifications in logs/checkpoints - Activity workers only - Opt-in (semantic_check=True) - No regression on atomicity tests - DOCUMENTATION: Update docs/HIGHWAY_WORKFLOW_ENGINE_BOOK.md - TESTING: Use ollama nomic-embed-text model