fix(#221): RTL/mirrored word collapse + table cluster sliding-window

[jacob-cotten]

Problem

issue-848.pdf (8 pages, Ghostscript-generated with alternating page orientations):

• Word accuracy on mirrored pages: ~0.6% (expected ≥90%). All 1506 chars collapsed into 1 word per page.
• Table detection: 0% across all pages.

Root Causes (confirmed against Python pdfplumber source)

1. upright computed incorrectly for mirrored text (char_extraction.rs)

Python: upright = trm[1] == 0 and trm[2] == 0 and trm[0] > 0

Rust had: trm.b.abs() < 1e-6 && trm.c.abs() < 1e-6 — missing the trm.a > 0 check.

issue-848 uses CTM a=-1 (horizontal mirror). Python produces upright=False; Rust produced upright=true. This single field mismatch caused all downstream logic to fail.

2. WordExtractor::extract() dispatched on char.direction, not char.upright (words.rs)

Python's char_begins_new_word() dispatches on upright, routing upright=False chars through TTB logic: interline axis = abs(curr.x0 - prev.x0) > x_tolerance. Adjacent mirrored chars differ by ~5-6pt in x0 → each char becomes its own word.

Rust dispatched on char.direction — mirrored chars have direction=Rtl but were routed to horizontal processing, gap formula gave 0 for touching chars, all chars merged into one word.

3. snap_group() used cluster_start comparison instead of sliding-window (table.rs)

Python's cluster_list uses: x <= last + tolerance where last = previous element.

issue-848 rect x0 values: [72.3, 74.8, 77.4, ..., 85.3] — spread=13pt, consecutive gaps ≤3pt. Old Rust compared each to cluster_start: element 10 (85.3) vs start (72.3) = 13pt → broke cluster early → no valid column boundaries → 0 tables detected.

4. cluster_words_to_edges() same bug (Stream strategy)

5. extract_text_for_cells_with_options() used caller-supplied direction for all cells

Should detect per-cell from actual char.upright / word.direction.

Fixes

File	Change
char_extraction.rs	upright requires trm.a > 0 — matches Python exactly
words.rs	extract() partitions on char.upright; make_word_with_direction() stamps Word.direction=Ttb for non-upright words
table.rs	snap_group() sliding-window (edges[i-1]); cluster_words_to_edges() same; extract_text_for_cells_with_options() per-cell orientation

Tests

• char_extraction.rs: not_upright_for_horizontal_mirror_text
• words.rs: 7 tests covering non-upright splitting, direction=Ttb invariant, regression guards
• table.rs: 3 tests — exact issue-848 x0 data collapse, genuine-gap split, Stream sliding-window
• crates/pdfplumber/tests/issue_848_accuracy.rs: 6 cross-validation tests (chars≥95%, words≥90%, tables≥80%, LTR regression guard)
• cross_validation.rs: cv_python_issue_848 promoted from cross_validate_ignored! to active

Test plan

• cargo test -p pdfplumber-parse — not_upright_for_horizontal_mirror_text passes
• cargo test -p pdfplumber-core — all 7 upright word tests + 3 table sliding-window tests pass
• cargo test -p pdfplumber --test issue_848_accuracy -- --nocapture — all 6 accuracy tests pass
• cargo test -p pdfplumber --test cross_validation — cv_python_issue_848 passes; no regressions on existing PDFs
• cargo fmt --check — clean

🤖 Generated with Claude Code

[jacob-cotten]

Closing — opened against wrong repo. Apologies for the noise.