A table is a table in every format

[WaylandYang]

Pitfall 26 of docs/design/prior-work.md, the open item left by #744: a table is a structure in every format, not only in HTML.

What lands

DOCX, spreadsheet and CSV tables go through the grid that #744 built for HTML. Until now a DOCX table came out as one paragraph per cell, a sheet as tab-joined lines and a CSV as pipe-joined lines with no header separator, so the block parser saw none of them as tables and the model read cells without their headings. utopia_ingest::table::render_grid takes rows of (text, span, indent) from any parser and runs the same classification and rendering: caption lines, stacked headers, sections folded into labels, symbols rejoined to figures.

• DOCX (parsers::docx_xml_to_text): w:tbl rows and cells are collected with w:gridSpan as the span and the first paragraph's w:ind as the indent; a vertically merged continuation cell stays empty; a table nested in a cell is that cell's text. Paragraphs before and after are untouched.
• Spreadsheets: each sheet is one grid under its # Sheet: heading; empty rows are skipped; a sheet that renders no table falls back to the tab-joined lines.
• CSV / TSV: the records are one grid.
• Two structural rules for grids that come with a header convention (first_is_header): the first row with two or more cells is the header even when its headings are years, and every later row with two cells or a lone figure is a record. A table with no figure anywhere (words only) is read as a header row followed by records, in every format.

The text layer of a PDF carries no table structure and is left as it is; a scanned PDF's tables already arrive from MinerU as pipe tables.

A runner to look before extracting: cargo run -p utopia-ingest --example render_doc -- <file> prints a file as the ingest layer reads it.

Checked

Unit tests: a DOCX table with a spanning header and "$" cells renders | Segment | Revenue | over | Cloud | $1,200 |; a nested table is its cell's text; a CSV whose headings are years keeps them as headings; a CSV of words is still a table. On three real files (a course-exemption roster .xlsx, a visa application form .docx, a simulation-export .csv): the sheet's title row becomes the caption and its header row the header, with the sequence-number rows as records; the CSV's comment lines become the caption and its 639 records a table; the form's tables render as tables, with the label-value interleaving of a form left as it is.

Not in this change

Extraction measurements on DOCX or spreadsheet corpora (none is in the benches yet; the NVDA cross-domain recut and the period-column check remain the open items of pitfall 26); PDF text-layer tables.

🤖 Generated with Claude Code