4-stage pipeline as Temporal workflow: Stage 1: ChunkAndEmbed — chunk text + generate embeddings Stage 2: ExtractEntities — LLM entity extraction with reflection Stage 3: ExtractFacts — pattern + LLM fact extraction Stage 4: DetectContradictions — pre-filter + LLM verification Stage 5: PersistSynthesis — save all results to DB Types: SynthesisInput, SynthesisResult, ExtractedEntity, ExtractedFact, ContradictionResult, PersistInput Retry: 3 attempts, exponential backoff (1s → 2s → 4s) Each stage fails independently with wrapped errors. Tests: 5 pass (success, contradictions, entity fail, chunk fail, fields) Build: clean, 36 packages pass
