Render EMF/WMF diagrams so Office figures are analyzed

[Paul Lizer (paullizer)]

Refs #1277

Problem

A real TIC 3.0 architecture document uploaded with Enhanced extraction turned on produced no image analysis and no log output at all, despite containing four diagrams.

Root cause: the images were EMF metafiles, and the embedded-image whitelist only accepted raster formats.

OFFICE_EMBEDDED_IMAGE_EXTENSIONS = ('.png', '.jpg', '.jpeg', '.bmp', '.tif', '.tiff', '.heif', '.heic')

Word stores pasted diagrams, SmartArt, Visio drawings, and charts as EMF, so this silently skipped what are often the most information-dense figures in a document. Worse, extraction returned before emitting any status message, so "no images in this file" and "four images found, all skipped" looked identical in the workspace log — which is what made the problem impossible to diagnose.

Why this isn't a one-line whitelist change

Neither Document Intelligence nor Content Understanding accepts a metafile, so EMF has to be rasterized first, and the usual options are all unavailable here:

• Pillow can't do it on Linux. Its metafile handler is gated on hasattr(Image.core, "drawwmf"), commented "windows only", because it is GDI-backed. Relying on it would work in local Windows development and silently fail in the container.
• An external converter isn't an option. The runtime is mcr.microsoft.com/azurelinux/distroless/python:3.12 — no shell, no package manager — so LibreOffice or Inkscape would mean copying a very large dependency tree into a distroless image.
• There was no shortcut. Dumping the EMF records showed roughly 2,000 genuine vector drawing records per diagram (Bezier paths, polygons, brushes, world transforms), with no embedded bitmap to extract and no EMF+ wrapper.

Approach

functions_emf_render.py renders EMF and WMF in-process using only PIL, io, and struct. No system packages, no external binaries, no platform branches, so behavior is identical in the container and locally. The container is unchanged.

It covers the record subset Office actually emits for diagrams: path construction, filled and stroked polygons, Bezier curves, rectangles and ellipses, pen and brush objects, world transforms, and text runs. Records outside that subset are skipped rather than failing the render, so output degrades in fidelity instead of disappearing.

Text drawn inside a metafile is extracted as well and attached to the chunk, so figure labels stay searchable even when the vision engine returns no description.

Verified against the real document

All four diagrams now rasterize, and I inspected the rendered output rather than trusting an exit code — they are recognizable Azure architecture diagrams with subscription boundaries, service icons, and connectors. Recovered labels include Event Hub Namespace, Log Analytics Workspace, Azure Firewall Table, SysLog Table, and CLAW Registered App.

diagnostics: {'candidates': 4, 'analyzed': 4, 'skipped': 0, 'skipped_reasons': {}}
image1.emf  766x605   rasterized=True  27 text labels
image2.emf  1285x831  rasterized=True  38 text labels
image3.emf  989x831   rasterized=True  27 text labels
image4.emf  989x831   rasterized=True  33 text labels

Two bugs surfaced during that verification and are fixed here: the EMF/WMF format sniff misfired because \x01\x00 is just the first half of the EMF record type (the reliable discriminator is the " EMF" signature at offset 40), and the EXTTEXTOUTW field offsets were wrong, which silently yielded zero text runs.

Observability

Embedded image processing is no longer silent. Extraction now reports how many image parts were found, how many were analyzed, and why any were skipped, with progress reported per image:

Analyzed 4 of 6 embedded image(s) with Content Understanding. Skipped: 2 too small.

The counts persist on the document as office_embedded_image_candidates, office_embedded_image_count, and office_embedded_image_skipped, so the outcome can be confirmed after processing finishes.

Testing

Six new tests covering metafile rasterization, format-detection failure reporting, and diagnostics that distinguish "no images" from "images found but all skipped" — 17/17 in the embedded image suite. The Content Understanding, legacy Word, Document Intelligence, and route policy suites all still pass.

Reviewer notes

• Rasterization is a description aid for search and citation, not a pixel-accurate GDI reimplementation. Gradients, complex clipping regions, and bitmap blits inside a metafile are not reproduced, and text is drawn with a default font rather than the original typeface. The separately extracted text keeps labels accurate regardless.
• The renderer is bounded against untrusted input: record counts, point counts, and output dimensions are all capped, and a malformed record is skipped rather than aborting the render.

Version bumped to 0.250.223.