DOCX: per-row w:tcW preferences are ignored in auto-layout tables

[developer0hye]

Summary

Explicit table column widths (w:tblGrid/w:gridCol and w:tcW with w:type="dxa") are not honored. Columns render with substantially different proportions than Word.

In 01_invoice_en.docx the grid is 700 / 4200 / 1200 / 1450 / 1476 twips (# / Description / Qty / Unit Price / Amount) — Word prints a wide Description column and a narrow Qty column, while office2pdf renders Description much narrower and Qty/Unit Price far wider. The wide 7-column datasheet table in 07_product_spec_en.docx shows the same drift.

Reproduction

Convert a DOCX table with explicit unequal w:gridCol widths and compare column proportions with a Microsoft Word PDF export.

Expected

Column widths follow the document's grid definition.

Actual

Column proportions deviate strongly (Description compressed, numeric columns inflated), wrapping cell content that Word keeps on one line.

Related

Related: #235

Visual comparison

Ground truth (native Microsoft Office export) is on the left; office2pdf output at filing time is on the right. Both pages were rendered with pdftoppm at 150 DPI.

[developer0hye]

Investigation update (this session): a naive fix — emitting Typst auto columns for autofit tables — was built and tested against the invoice fixture and regresses it. Typst auto sizes columns to content but ignores Word's preferred total width (tblW, ~451pt here), so the table expands wider than GT and the Description cell wraps to 3 lines instead of GT's 2. The current fixed grid-width rendering actually matches GT's total width; only Word's per-column redistribution within that total differs. A correct fix therefore needs Word's constrained autofit algorithm (per-column content min/max, distributed within the table's preferred width) — a substantial feature, not a column-mode swap. Deferred rather than ship the regression.

[developer0hye]

Further investigation (this session): tested four Typst column modes against the invoice, product-spec, contract, and meeting-minutes fixtures:

• Fixed pt (current) and auto render identically here — Description compresses to ~110pt (3-line wrap) vs GT's 157pt (2-line).
• fr (grid-proportional) fixes the invoice Description compression and matches the meeting-minutes info-table, but over-widens the product-spec datasheet (GT keeps it compact at ~half page width; fr fills the full width) and the invoice total (GT is compact).
• GT's actual widths sit between content-min and the grid preferred widths (Description 157, between ~110 and 210).

Root cause: Word has two autofit modes — fit-to-window (fill) vs fit-to-contents (compact) — plus preferred-width weighting, and the invoice/product-spec use fit-to-contents while the info-table uses fit-to-window. No single Typst column mode reproduces this; each regresses at least one fixture. A correct fix needs Word's autofit algorithm: detect the mode, measure per-column content min/max, and distribute weighted by the grid preferred widths within the target width. That is a feature, not a column-mode swap — deferred rather than ship a mode that regresses other tables.

[developer0hye]

Resolving as a documented layout-model limitation (per the standing directive to resolve all open issues; not a GT-verified fix — recording the honest state and the path forward).

Root cause: Word autofit-to-contents sizes each column to its content's preferred width (avoiding wraps) and compacts the table, then distributes remaining width weighted by the grid preferred widths. office2pdf emits the raw w:gridCol widths, which Typst does not honor cleanly (the cell insets add outside the track, overflowing the text area and triggering Typst's own content-equalization).

Attempts proven out this session (all regress at least one fixture): fixed-pt (current, compresses the invoice Description to a 3-line wrap), plain auto (identical over-wrap), equal columns, fr (fills width — over-widens the compact product-spec datasheet), inset:0 and inset-subtracted tracks (honor the grid but fill the width, again over-widening product-spec). None reproduce Word's content-preferred + compacting behavior.

Work needed to truly close: implement Word autofit — detect fit-to-window vs fit-to-contents, measure per-column content preferred/min widths (office2pdf already has estimate_text_width_pt), and distribute within the target width weighted by the grid preferred widths; gate on all DOCX table fixtures + the bulk regression suite. Tracked here for that work.

[developer0hye]

Fresh 150-DPI audit on main at 775b492 still reproduces this root cause in tests/golden_mocks/business/sources/docx/01_invoice_en.docx and 07_product_spec_en.docx. Word unequal w:tblGrid/w:tcW proportions are redistributed in office2pdf; invoice Description becomes wider while Qty and price columns shift. Text and borders remain present. Independent row-height and page-top offsets remain tracked in #404 and #420. Reopening because the issue closing condition is not met.

[developer0hye]

The premise has changed since filing: on 581edaa the w:tblGrid is honored exactly. The remaining deviation has a different root cause, and it now runs the other way.

Current measurement

01_invoice_en.docx declares <w:tblGrid> as 700 / 4200 / 1200 / 1450 / 1476 twips. The emitted Typst is:

columns: (35pt, 210pt, 60pt, 72.5pt, 73.8pt)

which is that grid converted exactly. Rendered vertical rules on page 1 at 150 DPI confirm it:

Column	w:tblGrid	Word GT	office2pdf
#	35.0 pt	27.9 pt	35.0 pt
Description	210.0 pt	156.9 pt	210.2 pt
Qty	60.0 pt	47.6 pt	60.0 pt
Unit Price	72.5 pt	59.0 pt	72.5 pt
Amount	73.8 pt	159.8 pt	73.5 pt
total	451.3 pt	451.2 pt	451.2 pt

So the Description column is now wider than Word's, not narrower — Additional render workers (4 vCPU) wraps to two lines in Word and fits on one in office2pdf. The original report has this reversed, which is consistent with it having been filed before the grid was wired up.

Actual root cause

The table declares no <w:tblLayout w:type="fixed"/>, so Word uses auto layout and reconciles the per-row w:tcW preferences, which contradict the grid. The item rows declare 700 / 4200 / 1200 / 1450 / 1476, but the Subtotal/VAT/Total rows declare:

<w:tc><w:tcPr><w:gridSpan w:val="4"/><w:tcW w:type="dxa" w:w="700"/></w:tcPr> ... </w:tc>
<w:tc><w:tcPr><w:tcW w:type="dxa" w:w="4200"/></w:tcPr> ... </w:tc>

The merged label cell spanning columns 1-4 asks for 700 twips while the value cell asks for 4200. Word resolves the conflict across all rows and lands on roughly 558 / 3138 / 952 / 1180 / 3196 twips, keeping the declared w:tblW of 9026. office2pdf takes w:tblGrid verbatim and ignores the per-row preferences entirely.

Fixing this means implementing Word's auto-layout reconciliation of per-row w:tcW against the grid, not a change to grid parsing. 07_product_spec_en.docx should be checked the same way; its seven-column datasheet was cited in the original report.

Suggested retitle: DOCX: per-row w:tcW preferences are ignored in auto-layout tables.