Repository navigation
Fullbleed 2.5.33: correct table caption ownership
Table captions now remain inside their owning logical Table in tagged PDF output, with one Caption containing its paragraph. Top and bottom captions preserve reading order through nested tables, page continuations, compiled rendering and separate records. Empty and hidden captions retain their layout without empty semantic tags. This resolves #98 for the covered saved-PDF cases.
Install with python -m pip install fullbleed==2.5.33 or fullbleed = "=2.5.33" in Cargo. This is the stable engine backport of the correction in 2.6.0a7. The new component authoring API remains in that opt-in preview. No runtime dependencies are added.
Verification includes 671 local Python tests without skips, 1,399 local native tests and 11 doctests. All 225 caption cases pass independent saved-PDF structure checks; page pixels and extracted text match the public 2.5.32 controls, and all 45 untagged controls retain identical PDF bytes. All 150 caption and 120 existing rich-table PDF/UA fixtures pass veraPDF 1.30.3 with clean stderr. The stable PDF/UA-2 negative controls fail as expected.
All 645 caption, rich-table and pagination PDFs match across the qualified references, Linux/Windows CI, the exact-tag release gate, a fresh public Python installation and a public Rust consumer. All 15 public wheels, the sdist and the crate were downloaded and checked against retained release artifacts or tagged source.
The independent 1,662-case CSS corpus retains 1,620 PASS, 24 FAIL and 18 REFERENCE-DISPUTED with unchanged scores, verdicts and page pixels. One repeated-header PDF gains a balanced Artifact wrapper and exactly matches the reviewed alpha correction. No references or tolerances changed. Machine checks do not establish human assistive-technology usability or general accessibility conformance; dedicated PDF/VT validation remains unconfigured.