DOCX visual audit against Microsoft Word ground truth

[developer0hye]

Scope

Compare public real-world DOCX fixtures against Microsoft Word PDF ground truth using paired image inspection.

Completed

• run-level page breaks collapse two-page documents to one page (fix: preserve DOCX run page breaks #193)
• cached fldSimple results duplicate dynamic page numbers (fix: discard cached DOCX simple field text #194)
• inline images from one paragraph stack instead of wrapping (fix: preserve DOCX inline image flow #197)
• EMF images can be dropped (fix: preserve EMF images in DOCX #196)
• WMF images can be dropped or render with inverted coordinates (fix: preserve polygon WMF images in DOCX #198)
• RTL table cell order is reversed (fix: preserve DOCX visual RTL table order #199)
• WPG text-box and grouped shape content can be dropped (DOCX: preserve WPG text-box and grouped shape content #200, fix(docx): preserve WPG grouped content #205)
• legacy w:pict and VML fallback content can be dropped (DOCX: preserve legacy w:pict and VML fallback content #201, fix(docx): preserve canvas image offsets #206)
• complex list restart values can diverge (DOCX: preserve complex list restart semantics #202, fix(docx): preserve list restart counters #207)
• complex table shading and merge fidelity can diverge (DOCX: preserve complex table shading and merge fidelity #203, fix(docx): preserve conditional table shading #208)
• positioned footer tabs and borders are not preserved (DOCX: preserve positioned footer tabs and borders #204, DOCX: preserve positioned footer tabs and borders #210)

Each child issue was implemented in its own worktree, branch, artifact directory, and focused PR. Dependency limitations found during the audit were fixed in the dependency fork and proposed upstream where applicable, including bokuweb/docx-rs#897 for paragraph-border reading.

Method

Microsoft Word is automated in an alert-free batch session, outputs are rendered at the same dimensions, and GT/Before/After images are reviewed visually. Pixel metrics are used only to rank candidates, not to determine correctness. New artifacts use progressive JPEG at quality 86 to keep repository size reasonable while preserving readable text and layout detail.

[developer0hye]

Two visual-audit fixes are now merged with durable Microsoft Word GT comparisons.

Run-level page break (#193)

Microsoft Word GT	Before	After

Cached simple field result (#194)

Microsoft Word GT	Before	After

The remaining checklist items stay open for separate focused fixes.

[developer0hye]

EMF content preservation is fixed in #196.

Microsoft Word GT	Before	After

The cow and wrench EMFs are now preserved through native EMF-to-SVG conversion. The WMF image and exact floating placement remain separate work.

[developer0hye]

Inline image flow is fixed in #197.

Microsoft Word GT	Before	After

Images from one Word paragraph now share one wrapping flow container. WMF support remains the next image task.

[developer0hye]

WMF preservation is fixed and merged in #198.

The real mixed-image DOCX now preserves all five images. The native WMF polygon converter also honors the document's negative Y extent, so Santa is upright rather than mirrored.

Microsoft Word GT	Before	After

Visual review: Santa/cat/cow are restored on the first row and wrench/clock wrap to the second row, matching Word's content and reading order. Minor font metrics and spacing remain renderer-dependent.

[developer0hye]

Fixed the DOCX visual RTL table ordering case in #199.

The converter now respects w:bidiVisual, reverses cells and column widths into Word's visual order, and preserves column spans, vertical merges, borders, and padding. The real LibreOffice table-rtl.docx fixture is now covered by a strict structure assertion on Linux, macOS, and Windows with Git LFS hydrated.

Microsoft Word GT	Before	After

The corrected output matches Word's cell order (B1 | A1, B2 | A2). The small remaining position difference is due to renderer margin/font metrics rather than table direction.