fix(pdf): preserve visual layout when extracting PDF text

[duncanmcqueen]

Problem

extractPdfText joined every pdfjs text item on a page with a blanket space. For real legal documents that mangles the text:

• wrapped clauses lose their line structure (the whole page becomes one flowing string)
• kerning fragments gain spurious spaces (constitut + e → constitut e)
• indentation and column layout (signature blocks, simple tables) is destroyed

Every downstream consumer — chat document context, citation quote matching, anything built from extracted PDF text — inherits that mangled text.

Fix

Rebuild each page's text from the positioned pdfjs items (transform, width, hasEOL) instead of the raw string list:

• Lines are reconstructed from y-coordinates and hasEOL markers, with a y-jump fallback for producers that don't set hasEOL reliably; lines are ordered top-to-bottom, items left-to-right.
• Word spacing is driven by measured x-gaps: kerning fragments join without a space, ordinary word gaps get exactly one, and only large gaps become multiple spaces (capped at 16) so signature blocks and simple tables keep their columns without exaggerating justified text.
• Paragraph breaks: a vertical gap above ~1.7× the line height becomes a blank line.
• Indentation is preserved relative to the page's left margin (capped at 24 spaces).

The page-marker format ([Page N]) and the unreadable-buffer `` fallback are unchanged.

Tests

New documentOps.test.ts with a mocked pdfjs covers: line reconstruction from positions, paragraph breaks, column-gap preservation, page markers/ordering, and the failure fallback. Full backend suite passes (607 tests), tsc --noEmit clean.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}