GWG 1.0 - CMYK Overprint Test

[erikbra]

Observed in artifacts/html-examples/ghent-output-x4-page/semantic-continuous/index.html without consulting the source PDF.

The font, vector, image, mask, and shading samples show visible failure crosses instead of the expected CMYK overprint results.

[erikbra]

This was very hard.

The bounded investigation confirmed that PdfLayoutExtractor.ResolveGraphicsColor converts CMYK to RGB before the HTML layer sees it. PdfLayoutColor retains neither the original colorants nor overprint state, while this GWG test spans text, vectors, images, masks, and shadings. The existing raster backends also do not implement overprint, so fallback rendering would remain incorrect.

A proper fix needs a cross-cutting colorant-aware compositing model rather than a local HTML adjustment. No speculative changes were made.

[erikbra]

Reassessed after unpdf#114. This remains very hard; no #679 patch was retained.

The new support is deliberately limited to repeated, near-identical, opaque DeviceCMYK vector contour clusters. Extending it to a normal one-to-one OPM1 path pair was tested during #114 and produced concrete regressions in this GWG 1.0 panel and Ghent page 4, so that extension was removed before merge.

The remaining #679 samples also cross representation boundaries:

• font: text/glyph paint does not retain a composable CMYK backdrop
• vector: ordinary overlaps need page-backdrop-aware composition, not geometry matching
• image and mask: colorants/overprint are known in some cases, but no shared per-component surface exists
• shading: layout gradients do not retain OP/OPM or source components

A correct fix still needs operation-ordered component surfaces spanning these paint kinds. #114 makes one narrow repeated-contour case possible but does not remove that architectural prerequisite, so the very-hard label remains appropriate.

[erikbra]

Reassessed again after the recent CMYK/overprint work (#121-#125).

I reimplemented the previously rejected immediate OPM1 path-pair composition and added exact-stroke handling, then ran the focused tests and regenerated the Ghent fixture locally. The target vector contour did compose, but the full six-page semantic HTML diff again showed unrelated valid artwork changing on page 4 (including an opaque red cross becoming white and ordinary borders being recolored). Tightening the stroke case removed the border regressions, but the one-to-one fill case still reproduced the page-4 regression documented above.

The experiment has been discarded with a clean worktree. #679 still needs an operation-ordered component surface spanning text, vectors, images/masks, and shadings; the recent narrow overprint improvements do not yet make this issue a safe bang-for-the-buck fix.