feat(pdf): draw annotations in the browser

[andiwand]

🤖 Generated with Claude Code

Phase 5 of docs/design/pdf-annotation.md. Stacked on #848. Third of four; bindings follow.

odr.annotation collects what a viewer marks up and hands back the payload PdfFile::annotate takes:

odr.annotation.setTool("highlight" | "underline" | "strikeOut" | "squiggly" | "ink" | null);
odr.annotation.setColor([1, 0.9, 0.2]);   // DeviceRGB
odr.annotation.setWidth(3);               // ink
odr.annotation.list(); undo(); remove(id); clear();
odr.annotation.getAnnotations();          // the json #848 consumes

Coordinates

Each page div now carries data-odr-page and data-odr-space — the latter is to_box⁻¹, which is what Transform2D::inverse was added for back in #843. A viewport point divides out the zoom (rect.width / offsetWidth, robust against the zoom script's CSS transform), converts CSS pixels to points, and goes through that matrix. The model keeps page-box points and maps to user space only on the way out, so there is one representation to reason about.

The bug worth recording

The highlight painted over the text. mix-blend-mode on a shape inside an <svg> composites against that svg's own canvas, not against the page behind it — so the wash covered the glyphs instead of letting them through, exactly the failure the whole /BM /Multiply design exists to avoid.

The blend has to sit on the overlay element, which means the washes need an overlay of their own (svg.an-m) separate from the marks drawn on top of the text (svg.an). Two overlays per page. Caught by eye in the browser, not by a test — which is why the harness below now checks which overlay a shape lands in.

Pointer handling

The overlay takes pointer events only while the ink tool is active. A text tool leaves the selection layer alone — that is what lets a viewer select text to highlight it at all, and it sidesteps the link-overlay-vs-selection conflict for the markup tools.

Verified

test/browser/annotation/ — a harness in the shape of test/browser/viewport/: serve lifts the script and style out of frontend.cpp so what runs is what ships, and tests.html stands in for a rendered view with two pages carrying real data-odr-space attributes. 22 checks pass, covering the quad's user-space mapping, Z-ordering, which overlay each kind lands in, page indices, ink through real PointerEvents, and undo/remove/clear.

One of those checks caught a mistake in the harness rather than the script: I had predicted the quad's position from the CSS top, but a text range's client rect follows the font's metrics, not the line box. It now measures the run's own rect independently of the script, per the convention the viewport README sets out.

End to end, against style-various-1.pdf: marked up five annotations in a real browser, took getAnnotations() verbatim, fed it to PdfFile::annotate, and the rendered PDF places every one exactly where the browser drew it.

379 tests pass; -Werror clean.

No changelog entry: #848 already covers the feature, and this adds no new public C++ surface.