Fullbleed 2.6.0a8 — component preview

This opt-in Python preview brings the stable 2.5.34 page-furniture and running-header corrections into the component package, alongside the preview's improved page-margin sizing and internal running-content layout.

python -m pip install --pre "fullbleed==2.6.0a8"

Ordinary pip install fullbleed continues to select stable 2.5.34. The component API under fullbleed.ui.experimental remains experimental. This release does not publish a new Rust crate or change the Node/C# packages.

Download the component showcase source ZIP to recreate Northstar and Common Ground through Python components and HTML/CSS, plus the authored accessible guide in both PDF/UA profiles. The extracted ZIP was executed against the public wheel and reproduced the verified PDFs. Components retain owned styles, named parts, literal class_name selectors, and HTML export with or without styles. Bootstrap components and exports remain static.

Example PDFs: Northstar invoice, Common Ground report, accessible guide / PDF/UA-1, accessible guide / PDF/UA-2. The invoice's two authoring paths produce identical PDF bytes; the report's paths produce identical page pixels. Each accessible guide profile reproduces identical bytes through its component and exported HTML/CSS paths and passes independent veraPDF validation.

Generated margin text and decoration now stay outside body reading order in tagged output (#103). Compiled reflow refreshes captured headers and footers when record values change, including bindings shared with body text (#104). Page-margin boxes resolve sizing and alignment together; internal deferred running content uses each final slot width, including named page sizes.

Verification retains 1,451 native tests, 11 documentation tests, 216 focused Python tests and 943 full Python tests. Windows/Linux CI, the exact-tag release, and a fresh public Windows installation reproduce all 489 saved-PDF regression cases. 326 matrix PDF/UA fixtures have successful independent validation evidence: 20 new validator runs and 306 reused reports tied to byte-identical saved PDFs. Both guide profiles have fresh independent validation. All 15 wheels and the source archive were anonymously downloaded and checked against the release artifact hashes.

The independent CSS corpus completed with 1,619 PASS, 25 FAIL and 18 REFERENCE-DISPUTED cases. Its workflow remains failing; all 5,489 saved outputs, scores and raw verdicts match the reviewed candidate. The complete corpus, validation evidence, publication evidence, verification manifest, and SHA-256 checksums are attached.

That corpus run predates one added web-demo preview PNG, which corrected the first CI attempt's fallback to the stable demo image. All runtime, build, comparator and reference inputs are unchanged, and rebuilding the final source produced an identical wheel. The initial failure, fixture repair and corpus-reuse receipt are retained; final-source Windows/Linux CI passes.

Measured-box fit does not establish complete painted-ink containment: shadows, transformed content and oversized descendants can still clip. Streaming output is not atomic. Changed bindings force reflow for documents with running content even when the header is static. Public PageBand and PageNumbers constructors remain design work. Rich-cell continuation with middle/bottom alignment remains unsupported. Dedicated PDF/VT validation and human assistive-technology testing remain unverified; fixture checks do not establish general CSS or accessibility conformance.

Source: ef359ac202ea521bf14c993bdb3caa068b913f0e. Qualification runbook.