extract_text: glyph-order transposition fragments on multi-column textbooks

[yfedoseev]

Summary

On multi-column academic textbooks with dense figures and tables, extract_text emits hundreds of tokens that look like the right letters in the wrong order — e.g. aaxons (axons), acclipmoshed (accomplished), accpamonies (accompanies), achrmaotic (achromatic), acNrouens (neurons), Actiatvion (Activation), actualyl (actually), aDostokysve (Dostoyevsky).

v0.3.25 (#316 rowspan reading-order fix) dramatically reduces the count — ~2,100 such fragments disappear from Principles of Neural Science alone. But residual transpositions remain, and the same class of bug surfaces on other dense textbooks in the corpus.

Reproduction

./target/release/examples/extract_text_simple \
  "pdfs_slow4/Principles of Neural Science, Sixth Edition -- Eric R. Kandel (editor); Steven Siegelbaum (editor); Sarah -- ( WeLib.org ).pdf" \
  | grep -oE '[A-Za-z]{6,}' | sort -u > /tmp/words.txt

# Real word counts are intact
grep -c '^axons$' /tmp/words.txt          # 1
grep -c '^neurons$' /tmp/words.txt        # 1
# But so are the garbled fragments
grep -cE '^(aaxons|acclipmoshed|accpamonies|achrmaotic)$' /tmp/words.txt   # 4

Root-cause hypothesis

These look like neighboring glyph objects being emitted in writing-order rather than visual-order within a single TJ or TJ-array, probably caused by:

• Per-glyph positioning where the content stream walks glyphs in a non-linear order (e.g. kerning-optimized output from a typesetter), and the reading-order pass groups them by y-band before fixing left-to-right sort within the band.
• Overlapping figure callouts / annotations whose glyph positions are within the y-band tolerance of body text.

Test case: pairs of letters consistently end up swapped (aa / tv / yl) rather than long runs being scrambled — suggests a sort comparator that's unstable or that compares x-coordinates with a too-loose tolerance.

Reference corpus

• pdfs_slow4/Principles of Neural Science (major reduction but residual in v0.3.25)
• pdfs_slow3/The Advantage (Lencioni) — similar glyph-order fragments, not measurably improved by v0.3.25
• Kevin Murphy Machine Learning: A Probabilistic Perspective — different manifestation (concatenated unspaced words like groupregularization, equationOptimizing)

Tested versions

• 0.3.23: ~2,100 fragments on Principles of Neural Science
• 0.3.25 (release/v0.3.25): ~few dozen residual fragments on Principles of Neural Science; pdfs_slow3 and Murphy textbook largely unchanged

Priority

Medium — degrades downstream NLP / search / LLM context quality but does not lose content.