[iAlturki]

With use_text_flow=True, word boundaries are tested using the upright char_dir even for rotated character groups. For example, a clockwise-rotated ALPHA BETA is returned as nine single-letter words, although the reported word direction is ttb.

Use get_char_dir() for each group's orientation, matching the existing line-based extraction path and word metadata. This also respects the legacy horizontal_ltr=False and vertical_ttb=False options. Character order from the PDF is preserved.

The two new regression tests use the existing issue-848.pdf fixture. They cover top-to-bottom and bottom-to-top rotated text, legacy right-to-left/bottom-to-top directions, extraction with and without extra_attrs=["fontname"], and preservation of returned source characters. Changelog and contributor credit are included as requested by CONTRIBUTING.md.

Validation on Linux, CPython 3.12.3:

• Original source with the new tests: 173 passed, 2 failed; both failures are the new regressions.
• Fixed source: 175 passed, no skips, plus HTML coverage generation.
• Black, isort, Flake8, strict mypy, and git diff --check pass.

Actual word boxes from the same synthetic input PDF:

Offline interactive demo and reproduction instructions include the generated PDF, exact word bounds, source commits, and capture script. The proof assets are on a separate branch; this PR contains the fix, tests, changelog, and contributor entry.

Prepared with AI assistance. Commit author: iAlturki.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.