#330 Sherlock Implementation: Agentic Reliability for Highway

PARENT ISSUE: Sherlock Implementation CONTEXT FROM PDF (arXiv:2511.00330): The paper proposes a framework for reliability-aware agentic workflows. It identifies that LLMs are probabilistic, yet workflows treat them as deterministic. Errors propagate and compound. KEY CONCEPTS: 1. Reflexion: Local verification loops (Generate -> Verify -> Retry). 2. Speculation: Optimistic execution of downstream tasks before verification completes. 3. Equivalence: Semantic rather than bitwise equality for caching. GOAL: Enhance Highway to support these patterns natively while maintaining bank-grade atomicity. ARCHITECTURE: - Use 'Activities' for long-running reflexive loops. - Use 'Sagas' for speculative rollback. - Use 'Embeddings' for semantic checkpoints. REQUIREMENTS: - DOCUMENTATION: Update docs/HIGHWAY_WORKFLOW_ENGINE_BOOK.md after implementation. - TESTING: Add comprehensive tests. Use ollama and deepseek-v3.1:671b-cloud model for testing (in tools/llm provider). - REGRESSION: No regression allowed. All @tests/integration/test_atomicity_baseline.py tests must pass.