Fullbleed 2.5.14 - Arabic text and styled bidi

Arabic combining marks now map separately from their base letters in PDF Unicode text, and internal direction wrappers no longer replace ordinary spaces. Styled inline layout resolves bidi levels across sibling styles and preserves nested isolation for dates and references.

This release also restores standalone unicode-bidi declarations, supplies HTML direction/bidi defaults, and respects author rules over those defaults, including in CSS layers.

The installed regression checks marked Arabic, styled siblings, isolated fields and narrow wrapping with independent PDF readers. Reader-specific reconstruction of mixed-direction lines still varies; this release makes no universal language-support or new PDF standards-conformance claim.

Python: python -m pip install --upgrade fullbleed==2.5.14

Rust: fullbleed = "=2.5.14"

Release notes and qualification runbook. Node and .NET engine upgrades are separate releases.

Try the Arabic and English invoice project, with included fonts and an example PDF. Its data, HTML and CSS can be customized locally; follow the guide to run it and customize the layout.

Validation includes 1,338 Rust tests, 562 local installed-wheel Python tests, and 38 RTL output checks on Windows and Linux. The eight RTL fixtures produced identical PDF bytes across those platforms. See the source CI and platform release checks.

The retained verification report includes public-package checksums, individual RTL and invoice checks, cross-platform PDF hashes, and the deployed guide/download checks.

The published 2.5.14 Linux wheel has now completed the pinned 1,662-fixture IronPress CSS corpus: 1,616 PASS, 27 FAIL, and 19 REFERENCE-DISPUTED, including 858 identical page-raster fingerprints. The upstream gate failed. These results use the corpus's unchanged fixtures, committed Chrome/WeasyPrint reference PDFs, and comparator; they do not establish universal CSS or browser parity. See the results, limitations, and reproduction guide and open failure tracker.

The complete corpus evidence bundle retains all 1,662 candidate PDFs, reference/candidate/diff images, raw reports, and provenance. A structured summary and SHA-256 checksums are also available. This is additional evidence for the existing release; it does not include an engine fix for the reported failures.