Limits · glyphline

Type3 fonts and text drawn as paths have no advance widths. Those items fall back to the text layer or the items method; scans without an OCR layer have no text at all.
Vertical writing is placed per glyph, but sentence outlines assume horizontal lines: a vertical sentence becomes one shape per column.
Hyphenated line breaks in PDFs are joined by a heuristic (letter, hyphen, line end, lowercase letter). A compound word that happens to break at its hyphen becomes two words.
Sentences come from Intl.Segmenter plus an abbreviation list (e.g., et al., Fig.) and layout breaks in PDFs (font size changes, block gaps, list markers, table cells). An unusual abbreviation can still split a sentence.
Special tokens split the way Intl.Segmenter splits them: 42 ± 3 is three words, CJK follows the segmenter's dictionary.
Letter-spaced text (tracked capitals in a heading) often comes out of pdf.js with a space between every letter, so each letter becomes a word.
Drop caps are usually a separate text item, often placed before the heading in the content stream, so the large initial and the rest of its word are two words.
HTML that changes needs a new source. Ranges are offsets into the text the source was built from.
Right-to-left text is positioned correctly; the cursor's stretch direction follows screen position, not reading order.