feat: select text on an original page, and act on it

[NeoVand]

What changed

Text on an original page can be selected now, and everything that begins with a selection works from it: play, explain, highlight, note.

Before this the page was a picture. Highlights and notes made in the reflowed view displayed on it, but nothing could be created there, because both explain and annotation start from a selection and there was nothing to select.

How

Rather than laying pdf.js's own text layer over the canvas and then working out what a selection means, the layer is built from the boxes the placement already used — one invisible span per word, positioned in PDF points and scaled with the sheet, so a zoom moves it instead of rebuilding it.

The payoff is that each word arrives already knowing its passage and its position within it. Reading a selection is then a matter of asking its endpoints what they are, and the result is the same shape the reflowed view produces — so playSelectedPassage, annotateSelection and openExplainBox take it unchanged.

Words a passage claimed only as a region are attributed to it too. A table row is placed by its cells rather than by the words of its spoken reading, so it carries rectangles but no per-word boxes; without this the title page of a paper — which is mostly tables — was the one part that could not be selected into anything. On "Attention Is All You Need" page 1 that is 351 of 391 words knowing their passage, now 389.

The selection actions and the explain panel became snippets rendered inside whichever view is showing, so they sit in that view's scroll and travel with the page.

Two small things worth knowing, both caught by testing rather than reasoning: the highlight overlay had to stop taking pointer events, or it stood between the words and the pointer trying to select them; and the spaces and newlines that make a copied selection read properly are the only children in normal flow, so at any inherited size they stacked up as a visible column of blank lines down the corner of the page.

Verification

• npm run lint
• npm run check
• npm run test:coverage — 529 unit tests; domain/pdf-layout.ts at 96.5%
• Relevant Chromium/visual journey — 47 e2e including the 12 committed visual baselines, unchanged

The new e2e drags across two words on a rendered page and asserts: the selection contains the expected word, the actions appear inside the page's own scroll, highlighting leaves a mark on the paper, and playing the selection starts playback.

Review checklist

• Privacy and network behavior are unchanged. No new data leaves the device; the word layer is built from geometry already computed locally.
• Storage/migration. No schema change. Annotations created from the page use the existing anchors and are indistinguishable from ones made in the reflowed view — they display in both.
• Keyboard, screen-reader, contrast, reduced-motion. The word layer is inert to assistive tech beyond being selectable text; the actions are the same buttons, with the same labels, as the reflowed view. Contrast and motion are unchanged.
• Visual changes include desktop screenshots — captured (selection with the action bar, a highlight created from the page) but not attached; the CLI cannot upload them.

Known limits

• A word selected inside a region-placed passage resolves to that passage's start rather than to the word — right for playing and annotating a table row, coarser than the per-word precision prose gets.
• Annotation markers (the margin buttons) still render only in the reflowed view; the highlight itself shows in both.

🤖 Generated with Claude Code