[crunz-ailab]

Summary

Use top / bottom when Table._get_header().row_has_bold() checks whether a cached character is inside a table row. The row rectangle uses top-down page coordinates, but this helper currently compares it with the character's CTM-transformed y0 / y1 values.

On an ordinary unrotated page, this can both miss a genuinely bold first row and treat unrelated, vertically mirrored bold text as belonging to that row. As a result, an otherwise correctly detected table can receive the wrong internal/external header.

This is a separate occurrence of the coordinate mismatch discussed in #5092 and PR #5096. Credit to Hao-ActuaViz for that report and BetterAndBetterII for the chars_in_rect() patch. This PR only addresses row_has_bold(); it neither duplicates that patch nor claims to resolve the open-table detection issue.

Change and scope

- if rect_in_rect((c["x0"], c["y0"], c["x1"], c["y1"]), bbox)
+ if rect_in_rect((c["x0"], c["top"], c["x1"], c["bottom"]), bbox)

The full-containment predicate, row bounds, and tolerances are unchanged. Nearby or partially overlapping characters do not become eligible simply because they are near a table. There are no changes to snapping, border completion, candidate grouping, chars_in_rect(), or table geometry.

Reproduction and tests

The added parameterized test constructs a closed 2-by-2 grid in memory, with no external PDF dependency. It covers four cases at two vertical positions:

• A bold first row with ordinary text above: keep the first row as the internal header.
• A nonbold first row with unrelated bold text at the mirrored vertical position: retain the actual external header.
• An ordinary external-header control.
• An internal-header control with no text above.

Each case asserts the table bbox, all four cell rectangles, extracted cell text, header names, and header.external.

For the first case, the cell text stays [['Header A', 'Header B'], ['value A', 'value B']]. Before the fix, the header is incorrectly ['Above A', 'Above B'] with external=True. After the fix, it is ['Header A', 'Header B'] with external=False.

Validation

Base: d2423b1367bcc9721ea7e994e710e5bc7ea91b09.

Local test scope	Unmodified source	This patch
Eight added parameterized cases	4 passed, 4 failed	8 passed
Entire tests/test_tables.py, including the new cases	50 passed, 4 failed	54 passed

Build limitation: these runs used current main Python sources with the existing PyMuPDF 2.0 compiled components / MuPDF 1.28.5, on macOS with Python 3.12.2 and pytest 7.4.4. They are not a full rebuild of current main (whose default MuPDF version is newer), nor a claim that the complete repository suite or upstream CI has passed. Layout and automatic dependency installation were disabled. A source-built CI run is still needed.

An additional local comparison of native find_tables(use_layout=False) covered the first 60 pages of three technical manuals (180 pages total). Table counts, table/cell geometry, and extracted cell text were unchanged; only header metadata changed on two pages. This is supplementary local regression evidence, not a guarantee of exhaustive document compatibility. The self-contained synthetic tests above provide the primary reproducible validation. No original manual pages or extracted text are redistributed.

Supporting material

The prepared evidence archive includes the synthetic PDF and its generation/reproduction script, actual before/after results, focused containment tests, the proposed patch, sanitized upstream test logs, an AST isolation audit, and the 180-page comparison summaries. The earlier PyMuPDF 2.0 evidence is clearly separated from the new main-source validation. No manufacturer PDF content is included.

AI tools assisted with investigation, test preparation and execution, and drafting this description. Reported results are from actual local runs; the runtime limitations above are explicit.

Attachments

• Evidence archive: scripts, PDF, results, tests, logs and audits
• Synthetic PDF
• Evidence usage notes

These files are stored on a separate evidence branch in my fork; they are not included in this PR diff. The links above pin the evidence commit.

Synthetic closed-grid example (page 2 of the PDF):

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[crunz-ailab]

I have read the CLA Document and I hereby sign the CLA

[crunz-ailab]

recheck

[veget-able]

Thanks for the fix. Using top/bottom makes the character bounds consistent with the row coordinates.

In my 168-page comparison, most header changes appeared helpful, and Table.extract() was unchanged. I'm happy to approve this coordinate correction.

[crunz-ailab]

@veget-able
Both matrix jobs failed with the same four ReadTimeout errors while downloading PDF test fixtures from GitHub. Each run had 595 passing tests. Could you please rerun the failed jobs? Thanks!