902 gold tables: analysis_run_llm_usage_by_model_name (5 rows, calls/input_tokens/output_tokens/cost_usd metrics); batches_by_* temporal/status/version aggregations (2–186 rows, total_rows/bytes/elapsed_ms); column_tags_by_* family/source/tag/temporal (1–2107 rows, confidence); columns_catalog_by_* arrow/duckdb/sqlite types & semantic roles (1–1273 rows, null/distinct counts, avg values); deleted_datasets_by_* source/temporal (1–9 rows, table/row counts); jobs_by_* status/phase/type/temporal (1–203 rows, progress_pct, error_count); llm_calls_by_* model/provider/purpose/cache_hit/error_type/temporal (2–2972 rows, tokens, cost, latency); metric_readings_by_* segment/period/measure (1–580 rows, numerator/denominator/value); neal_findings_by_* kind/severity/reducer