Audits a tabular dataset for data quality before any model is fit — per-column nulls, ranges, types, semantic class (ID / categorical / continuous / ordinal / text / datetime / boolean), cardinality alarms, outlier flags, and row-level integrity (duplicate rows, conflicting fact-pairs). Refuses to drop outliers without semantic context. Use whenever the user receives a new dataset, before fitting any model, when results suddenly look off, or when a dataset changed shape between runs.
