Keep nested records as nested records
Standard lists and structs can represent information that does not fit one scalar spreadsheet cell. The workbench keeps their structure in JSON. In CSV and Excel, a nested value is serialized as JSON text inside a cell; it is not expanded into every possible combination of child elements.
An empty list is different from a null list. A struct containing a null child is different from an absent struct. The precision example contains these cases, and the typed JSON output preserves the distinctions. For binary columns, hex text preserves bytes without guessing that arbitrary data is UTF-8.
Compare a meaningful set of values
Check more than the first cell. Compare a large identifier, a negative decimal, a null, an empty string, a nested record and a row on either side of a row-group boundary. Confirm that the source-row numbers and matched count agree with the requested window and filter.
The schema manifest records the selected fields and their original annotations. Keep it with CSV, especially when downstream software does its own type inference. A local conversion is not a repair operation: missing or corrupt pages are not recreated, and unsupported types are not silently rewritten to resemble supported values.
Know when an export is unavailable
The file and scan limits are visible before upload. This release refuses encrypted or external columns, INT96 timestamps, maps, variants and unsupported list layouts. Dotted or reserved field names are also refused to avoid ambiguous field paths or unsafe object-key handling. No remote resource is loaded from file metadata.
Excel cannot hold arbitrarily long cell text. If any selected value exceeds 32,767 characters, the workbench omits the workbook and explains why; complete CSV and typed JSON remain available within the output budget. An explicit unavailable format is preferable to a workbook that quietly loses part of a value.