This path flags pipeline_operator: venue categorization runs after extraction; inverted indexes / embeddings / competitor finalize always afterward. Deep-research disk export and Phase 2.7 research_priority / account_signals builds are omitted (use research jobs); ML inference parquet export stays off unless you POST auto_ml_intelligence: true without pipeline_operator from the API. ML Enrichment + Abstracts are forced on server-side so Activities/Campaigns and abstract-gated venue tagging always participate. When categorize finds zero venues queued for LLM batches (same junk/title rules as production), processed_records is touched so the Pipeline table “Updated” clock reflects that verify pass.