v2.142.0

Feature

• afp: Treat a move to a new baseline as a line boundary (#4962) (af3e144)

Fix

• cli: Exit with code 1 when a document fails to convert (#4888) (787ad5e)
• nemotron-parse: Bound work on untrusted model output (#4923) (3b0f25f)
• Align transformer engine input dtype with model weights (fixes #3110) (#4158) (a401629)
• asr: End merged sentences at full-width CJK punctuation (#4836) (d18c25b)
• Adapt to docling-core 2.103.1 image and HTML changes (4e6a615)
• deps: Require docling-core 2.103.1 (197a65c)
• vlm: Keep every markup heading depth distinct across VLM mappers (73b7c97)
• vlm: Share the Nemotron 3.5 formula delimiter normalizer (d2b076a)
• Normalize Nemotron formula delimiters before export (91332c9)
• Preserve VLM section heading depth in exports (fec45d3)
• Preserve multiline DeepSeek and Unlimited OCR blocks (62eeaee)
• ebcdic: Report a record length shorter than its layout (#4961) (6470556)
• odt: Extract page headers and footers (#4960) (e3d4743)
• Handle truncated PNG chunks in DOCX pictures (#4946) (10054ec)
• latex: Backend updates for better parsing of arxiv papers (#4817) (29caea8)
• otsl: Keep cells when an unclosed cell precedes a closed one (#4916) (7b9a6b1)
• Support Pillow 13 and integer PNG images (#4929) (ed66a3c)
• datamodel: Read one content-sniff window for a path and a stream (#4924) (ffa6a84)

Documentation

• vlm: Correct the heading rule docs and comments (91a6439)

Performance

• Reduce scalar reads in image classification (#4918) (dae5636)
• ocr: Avoid temporary PNG files in macOS OCR (#4919) (44c4096)
• ocr: Convert PIL crops with numpy.asarray (#4920) (de7a4ea)