fix(flash): measure word gaps from each char code's font width

[rejojer]

Flash split words such as "infe rence" and "Va riational" in page text and in titles, so a chapter heading like "Monte Carlo inference" no longer matched its bookmark.

Cause

The gap before a character was measured from the previous glyph's ink box whenever the ink reached past the glyph's advance, as the tail of a bold-italic "f" does. The ligature rule then held the pen at that ink edge, so the next letter looked like the start of a new word.

Fix

Measure from the pen position, as the extractor Flash was ported from does: the font's width for the character code the content stream draws. The existing code walk already pairs each character with that code, so it now also reads the width from the font dictionary (/Widths, else /MissingWidth; Identity-H /W, else /DW). PDFium's GetGlyphWidth looks widths up by unicode and can land on another glyph's width, for example a summation sign drawn with a letter code.

Only text whose pen moves along +x uses the code width: upright, with a positive font size, read left to right.

• Rotated or mirrored text keeps the old measurement, since its pen does not move along +x.
• Right-to-left characters keep it too. PDFium returns them in logical order, so the code walk pairs each one with its mirror's code.
• Type 3 fonts keep it as well: their exact widths split headings ("Typ es") that the box end keeps whole.
• So do characters the walk can't pair, and a character that a later walk re-pairs with a font that has no widths.

When PDFium drops a glyph that overlaps the one before it, such as the second "f" of an italic "ff" drawn one glyph per show op, Flash re-emits it. The re-emitted glyph now starts at the previous glyph's pen end rather than its ink edge. When no later glyph gives its advance, it advances by its own code width. Before, a 10-K read "the ef fects of".

Results

Measured on 16 documents (3,579 pages) against the original extractor's text items:

• 4,744 word gaps fixed. 7 got worse, all in math or references, where the gap sits exactly on the 0.1 em threshold.
• Outlines change in 2 documents, both for the better: titles get their whole words back, and a formula no longer becomes a heading.
• CPU +4% on a 1,098-page book.

Rotated text in 34 other documents and 5,380 generated lines mixing Hebrew with Latin, punctuation and digits read exactly as before.

Tests

• test_ink_past_the_advance_does_not_split_the_word: a narrow f whose ink overlaps the next letter. It reads "infe rence" before the fix and "inference" after.
• test_dropped_glyph_after_ink_past_the_advance_does_not_split_the_word: an "f" painted where the word's second "f" lands makes PDFium drop that glyph. Without the fix the word reads "ef fects".
• test_dropped_glyph_with_no_next_survivor_advances_by_its_code_width: a dropped glyph with no later glyph starts at the previous glyph's pen end and advances by its own code width.
• test_rotated_text_ignores_code_widths: "O'Brien rock'n'roll" rotated 270°, which code widths would read as "O'B rien rock'n 'r oll".
• test_rtl_chars_ignore_code_widths: a Latin line with two Hebrew letters keeps its order and gaps.
• test_composite_font_widths_read_both_w_forms: both /W forms, /DW, and a bounded range.
• build_pdf in tests/conftest.py also accepts a page's raw content stream.
• Full suite: 844 passed. The flash tests also pass on Linux and on pypdfium2 4.30.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
🔒 Security Review	✅ Completed 2026-10-07T09:22:41.632069Z	c155a32	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.