Keep the PDF's advances when hayro stretches a substitute glyph #656

Context

DESIGN.md section 14: pyropdf is compared with frozen pylopdf. #609 moved pyropdf to hayro 0.9, after 0.2.1, which is on hayro 0.8. hayro 0.8 stretched the outline of a substitute glyph to the width that the PDF expects (stretch_glyph in hayro-interpret/src/font/). hayro 0.9 puts that stretch in the glyph's transform instead (Affine::scale_non_uniform(font.glyph_stretch(glyph_id, code), 1.0) in hayro-interpret/src/interpret/text.rs).

Text extraction takes a glyph's advance through that transform (combined * Point::new(adv, 0.0) in collect_glyph in crates/pyropdf-engine/src/extract/text.rs, and likewise in record_text_item in model.rs and widen_holder). So a font that a document does not embed, drawn in a Standard 14 substitute whose widths differ from the PDF's /Widths, gets each advance scaled by that glyph's stretch. The corpus of evidence/phase1-parity has no such page, so the parity report does not show it.

Found while working on #624. A page in /BaseFont /Verdana, not embedded, with /Widths of 700 throughout, at 20 points, gives these page.words() boxes:

words(exact_bboxes=True) matches, since ink goes through the same transform on purpose, but chars without ink in layout(chars=True, exact_bboxes=True) do not: a space ends at 125.25 and not at 104.0. Word gaps also read these advances, so a stretch below 1 can split a word and one above 1 can join two.

0.2.2 is meant to move no text (#622, #636), so this is a regression of that release if it ships as it is. The owner decides whether it goes into 0.2.2.

Goal

A glyph's advance in text extraction is the PDF's width, whatever stretch hayro applies to draw it, as on hayro 0.8 and in frozen pylopdf.

Scope

In: the advances of page.text, words, layout, chars, search, selection, and tables, and of the text runs of page.content(), for fonts that a document does not embed.

Out: the ink boxes, which should keep the stretch, because a rendering draws it. Exact boxes of CJK fonts on fallback faces, which #624 handles by taking only ink from the faces.

Paths

crates/pyropdf-engine/src/extract/, the tests of text boxes, and a new page in evidence/phase1-parity/ if the fixture joins the corpus. crates/pyropdf-engine/Cargo.toml and tests/test_crate_dependencies.py if the fix depends on hayro-font directly.

Acceptance

• On a page in a non-embedded font whose /Widths differ from its substitute's, words(), layout(chars=True), and search() match frozen pylopdf.
• The advances of page.content()'s text runs on that page are the PDF's widths.
• Ink boxes on that page do not change.
• evidence/phase1-parity measures such a page, so that this cannot come back unseen.

Notes

• The stretch is not reachable from hayro_interpret's public API: OutlineGlyph and PositionedGlyph do not expose it. hayro_font::Font::glyph_stretch and FontCache are public in hayro-font, but hayro_interpret::font re-exports them as pub(crate), so the options are a direct dependency on hayro-font that loads a font again for runs in a font without an embedded program, or an upstream accessor on the glyph, which would then wait on hayro or on the fork that Build against released hayro again: remove the [patch.crates-io] of LaurenzV/hayro#1734 #603 removes.
• GlyphRun::font_dict() tells which font a run uses, which is enough to skip embedded fonts: hayro stretches only a substitute.
• The reproduction is a Type1 or TrueType font dictionary with /FontDescriptor and no FontFile*, and /Widths far from Helvetica's.