PDF Craft v2.6.1

PDF Craft 2.6.1 fixes cross-page paragraphs interrupted by figures or tables,
adds last-resort recovery for inconsistent extraction results, and prevents
contradictory OCR geometry from erasing unrelated PDF content. It also fixes
repeated XML-fill failures when translation turns LaTeX into natural text.

Public APIs and the PCEX v4 archive format are unchanged from 2.6.0.

Extraction and recovery

• Keep images and tables between the text fragments of a cross-page paragraph,
  including during footnote refinement and reconstruction. Independent display
  formulas remain paragraph boundaries.
• When permitted by the error policy, unusable chapter analysis can recover
  affected pages as source-page images. If rendering fails, a separate policy
  decision permits explicit page placeholders. Recovery also covers extraction
  without LLM refinement.
• Recovery includes complete cross-page paragraphs and related footnotes and
  references, including intermediate pages containing only anchored assets.
  It leaves OCR caches and already delivered previews unchanged.
• Accepting the traditional algorithmic result does not automatically authorize
  page-image recovery. Decision callbacks receive the proposed recovery and
  affected pages. If every primary page ultimately degrades because of errors,
  extraction still fails instead of reporting success.

PDF writeback protection

• Detect text source boxes that substantially cover unrelated figures or text
  before erasing them. With policy approval, retain the actual original PDF
  pages, including pages linked by complete cross-page paragraphs, while other
  pages continue processing.
• Skipped replacements do not emit fabricated translated layout previews. Strict
  error policies still raise, and recovery of every primary writeback target
  still makes the task fail.
• Fix a first-item translation error when page furniture supplies a
  pretranslation mapping.

These extraction and writeback fixes are included in
#525.

Formula translation and XML fill

• Treat temporary formula elements as ordinary inline text slots during fill.
  A valid translation such as “16 times” no longer has to reproduce the source
  LaTeX spelling to pass validation.
• Keep strict structure, ID and full-paragraph text validation. Restore original
  source formulas only after the fill loop accepts the response, without
  restarting the loop because their representation differs from the translation.
• Add deterministic regressions for the three reported paragraphs and a reusable
  real-model gate requiring three consecutive successful runs at temperature
  0.7 and top_p 1.0, without counting fallback as success.

The English and Chinese error-handling and PDF-translation guides describe the
recovery decisions and formula-restoration behavior.

Additional fill edge cases remain tracked in
#526 and
#527.

Full Changelog: v2.6.0...v2.6.1