902 gold tables: analysis_run_llm_usage_by_model_name (5 rows, calls/tokens/cost metrics), batches_by_* temporal/status aggregations (2–186 rows), column_tags_by_* family/source/tag groupings (2–2107 rows), columns_catalog_by_* type/kind/role dimensions (1–1273 rows), deleted_datasets_by_* source/date (1–9 rows), jobs_by_* status/phase/type/date (1–203 rows), llm_calls_by_* model/provider/purpose/cache (2–2972 rows), metric_readings_by_* segment/period (1–580 rows), metrics_by_* direction/origin/status (1–2 rows), neal_findings_by_* kind/severity/reducer (1–258 rows), neal_probe_runs_by_* probe_kind/status/refusal_code (1–185 rows), neal_readings_by_* group_axes/measure (1–60 rows), nia_message_ll