PDF Craft v2.3.0

What's Changed

Upgrade PCEX chapters to a v3 FlowItem-based document flow. A text flow now
preserves its source fragments together with image/table assets anchored inside
that flow, while display formulas remain explicit reading-flow boundaries.
Canonical writers now emit PCEX format v3. Readers retain migration support for
v1 and v2 archives.
Render anchored images and tables in Markdown and EPUB in reading order. EPUB
uses a conservative responsive float only when source geometry provides strong
side-by-side evidence; otherwise assets remain normal block content.
Translate image/table title, content, and caption fields through a separate
anchored-content stage, so they do not pollute NarrativeFlow translation context.

Add ExtractionOptions.includes_furniture (default False) to opt into
extraction of native PDF page furniture into furnitures.xml. It identifies
recurring Universal and SameSide patterns, retains unmatched page-local
sections, and supports common running heads, footers, and folios.
Add PDFCraft.translate_furnitures() and its FurnitureTransformer contract
as a separate page-furniture translation stage. Translation coverage is
recorded in translation.xml, allowing translated furniture to be patched
back onto PDF pages without entering the main narrative translation flow.
Connect printed table-of-contents pages, body headings, and matching furniture
sections through TOC identities so repeated chapter titles can share a
translation.

Refine two-stage PDF text fitting, including normalization for isolated
single-line source boxes whose OCR height is not a reliable font-size signal.
Keep anchored image/table positions out of PDF text fitting: source text
fragments in one TextFlowItem remain one continuous PDF text flow while the
source artwork stays in the visual base.
Introduce read-only logical text projections for PCEX flows, shared by
Narrative translation, PDF patching, Markdown, EPUB, and anchored-content
context.
Preserve the canonical first-stage translation through XML fill. Fragment and
anchor slots can no longer introduce visible spaces or split continuous text in
PDF output, including cross-page fragment boundaries.

Add ExtractionOptions.extract_book_metadata (default False) and its
required, independent metadata_llm to enable front-page OCR metadata
extraction for PDF-to-EPUB and reusable PCEX workflows. The metadata agent
begins with three OCR pages and may inspect up to twelve front pages when more
evidence is needed.
Validate metadata JSON through the existing repair loop and retain only
page-evidenced fields. Native PDF Info is used only to fill OCR-missing
metadata.
Extend PCEX manifests with richer bibliographic metadata, including original
names, publication details, subjects, and rights.

Add PDFCraft.translate_anchored_contents() and its
AnchoredContentTransformer contract for a separate image/table text
translation stage.
PCEX v3 archives can contain optional furnitures.xml and translation.xml
members. The latter records translated / preserved coverage for Narrative,
furniture, and extracted image/table text, so a PDF patch only replaces text
that has actually been translated.
The expanded manifest.json document object now records original_title,
structured author original_name / nationality, publication_date,
edition, subjects, and rights.

Remove callback-based PDF translation (Callable[[str], str]). Translate a
PCEX with a structured chapter transformer, then patch the resulting extraction
with PDFCraft.patch_pdf_with_extraction(); PDFCraft.translate_pdf() now
accepts only a structured chapter transformer.

Full Changelog: v2.2.1...v2.3.0