Ingest and table detection

Spreadsheets are read as grids (SheetJS), not as text. On each sheet the engine finds table regions: blocks of non-empty cells with a header row, including merged or multi-row headers, several tables on one sheet, subtotal and total rows (marked _is_total, never summed), and transposed tables. Each table becomes a DuckDB table with typed columns and a _row number that points back to the sheet row — which is what citations use.

CSV files are sniffed for delimiter, encoding and quoting. Documents are split into sections that keep their page, heading path and character span: PDF text layers (unpdf), DOCX (mammoth), emails (mailparser, with headers), Markdown and text. Scanned or hard PDFs can go through an optional Docling container.

Limits apply per file and per job (size, sheets, cells, a ZIP-bomb check), and every problem is reported as an ingest issue rather than silently dropped.

See docs/specs/ingestion-and-profiling.md for the full rules.