[tannevaled]

What

A vector backend for the paint layer, and the API to drive it page by page:

• paint.Canvas — a small vector surface (Save/Restore, Clip(path, evenOdd), Fill(path, colour, evenOdd), Shade(Shading), Image(img, rect), Glyphs(*GlyphRun)), with Path (lines, cubics, RoundRect, Ellipse).
• paint.PaintVector(c, box, fonts, imgs, bgImgs, page) VectorStats — walks the box tree in paintBoxContent's order with its geometry and clipping, emitting paths, shadings, images and glyph runs.
• engine.(*Engine).LayoutHTML / LayoutDocument → *PageLayout with Paint() (the image RenderHTML returns, byte-identical — tested), PaintVector(c), Links(), Info(), Page().
• pdfcanvas — the paint.Canvas for a go-pdfkit page (needs go-pdfkit v0.15.0: ShowGlyphs, NewFont, radial shadings).
• css.GradientSampler.Geometry() — the resolved axis/ellipse and stops, so a backend can emit a shading instead of sampling.

Design note — why this shape

Read first. paint.go paints straight into an *image.RGBA (go-widgets/painter for rects, coverage masks for glyphs, per-pixel loops for gradients/shadows/tiles); there is no display list or painter interface. Two options:

1. Refactor paint.go behind an interface so raster and vector share one walker. It touches ~1,800 lines that other sessions edit daily, and every raster primitive (per-pixel erf shadows, group buffers, filter chains) would need a vector twin or an escape hatch inside the abstraction — a large, risky diff for the raster path, which must stay byte-identical.
2. A second walker over the same layout.Box tree (chosen). Layout is shared — positions, line breaks, fonts and image sizes come from the one layout pass — so the vector page has the raster page's layout by construction; only painting is duplicated, and it mirrors paintBoxContent step for step (shadows, background, layers, inset shadows, borders, overflow/clip clipping, marker, inline fragments, items, children). The raster path is untouched.

What is vector: backgrounds and borders (rounded: a ring between two rounded rects, even-odd), sharp box-shadows (clip-excluded / even-odd), linear and radial gradients (type 2 / type 3 shadings; a CSS gradient is split into runs of constant opacity — the uncover theme's hard stop to transparent is vector), text as positioned glyphs of the faces the layout measured with (exact advances, as Fonts.Measure; fallback-face runs; rlig ligatures carry their whole text for /ToUnicode), <img> and background images placed in their boxes (the source bitmap, so identical images dedupe across pages), list markers, underlines, overflow and clip: rect() clipping.

Per-element raster fallback for what has no vector form here: a subtree under opacity < 1, filter, mask-image, transform: rotate() or backdrop-filter; a blurred box-shadow; a gradient whose opacity fades between stops; a form control; a background image tiled > 1024 times. The raster painter paints just that piece into a page-sized scratch buffer (bounded: 64 Mpx), cropped to its ink and placed as an image. The text of a rasterised subtree is laid over its bitmap invisibly (text render mode 3), so it stays selectable. VectorStats counts it all.

Alternative considered: go-pdfkit/html2pdf already drives the layout tree into pdfkit, but it is a paginating document exporter with a reduced painter (no radius, gradients, shadows, @font-face, opacity); a full-fidelity per-page painter belongs next to the raster painter it mirrors, and html2pdf can adopt PaintVector later.

Bibliography (opened)

ISO 32000-1:2008 (Adobe's PDF32000_2008.pdf): §8.5.4 Clipping Path Operators (W/W* n), §8.7.4.5.3 Type 2 (axial) and §8.7.4.5.4 Type 3 (radial) shadings, §9.3.6 Text Rendering Mode (mode 3), §9.4.3 Text-Showing Operators (TJ), §9.10.3 ToUnicode CMaps, §11.6.4.4 Constant Shape and Opacity, §11.6.5.3 Soft-Mask Images. Engine: paint/paint.go, paint/fonts.go, css/gradient_sample.go, engine.go (renderCore, newCanvas), linkmap.go.

Verification

• paint/vector_test.go has a reference rasteriser for the Canvas (point-sampled winding, shadings, images, glyph masks) and checks pixel parity against PaintFull on every feature (backgrounds, borders, rounded rings, 13 gradient shapes, background tiles, 8 shadow kinds, markers, images with object-fit and filters, groups, overflow and clip): ≤ 2–4 % of pixels off by > 64 (antialiasing and sub-pixel glyph positions). paint coverage stays 100 %, css 99.5 %.
• pdfcanvas tests read the PDF back with rsc.io/pdf: Type0 fonts with /ToUnicode (é included), axial and radial shadings, XObjects, 3 Tr for the ghost text.
• In go-slides (poppler pdftoppm at 96 dpi vs the raster render, per slide): 100-slide deck 0.46 % mean pixels off (worst slide 1.2 %), text extracted by pdftotext; 758 KiB against 10 MiB raster and marp-cli's 837 KiB.

MINOR: new API → next minor tag after merge.

🤖 Generated with Claude Code