dataset-triagelisted

Find mechanical data risks before downstream work. Report evidence. Do not mutate source data. ## Scope Use for row-and-column data. Skip images, free text corpora, audio, and unstructured documents. Triage detects shape, parsing, missingness, duplicates, key, cardinality, range, and format risks. It does not prove domain correctness. ## Inputs - Source path, query, table, or DataFrame. - Expected grain and row meaning, when known. - Expected keys, date range, units, and critical fields. - Safe sample limits and access constraints. Human owner confirms business rules, valid ranges, units, and key meaning. ## Workflow ### Step 1: Acquire Safely Use existing loader and dependencies. Record source, sheet/query, row filter, sample method, and load errors. - File: inspect extension, delimiter, encoding, sheet, header, and totals rows. - SQL/KQL: start with bounded recent window when timestamp exists. Avoid full-table pull. - In-memory frame: preserve original object; inspect copy when conversion needed. If sample lacks enough rows or classes, widen deliberately. Never claim sample represents full table without evidence. ### Step 2: Pin Structure Report row count, column count, column names, dtypes, first rows, last rows, and index shape. Flag: - Empty dataset. - Duplicate or blank column names. - Header or totals row parsed as data. - Numbers or dates stored as text. - Shifted or malformed rows. ### Step 3: Run Core Checks For pandas, adapt this compact baseline: `