• Fixed: an FB2 file whose XML prolog declares utf-8 while its bytes are legacy Cyrillic — or which declares no encoding at all — parsed as ParseOk with the title, chapter titles and whole body replaced by U+FFFD. A lenient UTF-8 decode that comes back mostly replacement characters is now taken as evidence against the declaration, and the bytes are retried as windows-1251 and koi8-r. Mislabelled FB2 is common in public catalogues, so this turns a class of unreadable books readable; genuine UTF-8 with a damaged run is unaffected.

Initial release.

• EPUB and FB2 parsing into one shared document model: BookDocument → Chapter → ParagraphBlock / HeadingBlock / ImageBlock, down to lazy Sentence and Word spans.
• bookParserFor entry point with magic-byte format detection and transparent .fb2.zip unwrapping; the archive layer (inspectBookArchive) is exported for the ambiguous cases.
• A cheap parseMetadata path for both formats — the FB2 side streams events instead of building a DOM, so neither format reads chapter content to answer with a title, authors, language, and cover.
• Expected failures returned as ParseResult (corrupt, unsupportedFormat, encoding, emptyDocument, drmProtected), never thrown.
• Script-driven rule-based segmentation behind the replaceable TextSegmenter port, with a per-writing-system support boundary.
• Language normalization against the whole of ISO-639-1, including an ISO-639-2 mapping.
• Opt-in JSON serialization (package:ebook_parser/serialization.dart) versioned by kBookDocumentSchemaVersion, with image bytes handed back by reference.