v0.9.6

• Fixed silent truncation of extracted text at Tika's 500,000-character default limit -
  text_extract now allows up to 10,000,000 characters (a fixed constant, deliberately not
  configurable so independent parties derive identical ISCCs) and logs a warning when the cap is
  reached; Text-Code and granular text features change for documents longer than 500,000
  characters
• Updated iscc-lib to 0.7.0 (now the minimum required version), which fixes Image-Codes longer
  than 64 bits to follow ISO 24138:2024 - Image-Code units at 96, 128 or 256 bits change beyond
  the first 64 bits and should be regenerated; 64-bit Image-Codes and ISCC-CODEs are unaffected
• Updated locked dependency versions (notably Pillow 13.0.0, exiv2 0.19.2, xxhash 4.0.1,
  pypdf 6.21.1, pypdfium2 5.14.0, resvg-py 0.5.0, numpy 2.5.4, bitarray 3.12.2) - with
  Pillow 13 EPUB cover thumbnails come out slightly different, and with exiv2 0.19.2 files written
  by image_meta_embed differ at the byte level (embedded metadata is unchanged)
• Hardened EPUB XML parsing with an explicit parser that disables entity resolution and network
  access (#156)
• Fixed EXIF XPTitle (and other Windows XP* tags) being extracted as a raw byte listing instead
  of text - changes name and the Meta-Code for affected images (#163)
• Fixed XMP struct containers being extracted as the placeholder type="Struct" (for example as
  name) - they are now skipped so the next mapped field or the file name applies (#165)
• Fixed code_meta crashing with UnicodeEncodeError on image metadata that is not valid UTF-8,
  such as legacy IPTC text - such values are now decoded as Latin-1, the IPTC default (#164)
• Added support for M4A audio (audio/mp4, with audio/x-m4a normalized to it) - .m4a files were
  sniffed as video/mp4 and failed with "No video stream detected"; ISCC metadata is embedded in
  M4A files as ISCC_* tags because TagLib drops MP4 tag names that contain a colon (#166)
• Changed text_features to drop a final chunk shorter than the CDC minimum chunk size (128
  characters at defaults) unless it is the only chunk - its simprint carried too little evidence
  and matched unrelated documents; retained simprints, offsets and sizes are unchanged, but
  persisted granular text features lose such tail entries (#162)
• Fixed per-call text_avg_chunk_size and byte_offsets options being ignored for granular text
  features - text_features read the global setting and code_text did not forward its options

Changes to generated output

Re-generate stored results for affected assets if you compare them with codes from earlier
releases:

• Text-Code and granular text features of non-PDF documents longer than 500,000 characters
• Image-Code units at 96, 128 or 256 bits (bits beyond the first 64)
• Meta-Code of images with EXIF XP* titles, XMP struct values, or non-UTF-8 IPTC/EXIF text
• Granular text features: undersized final chunks are no longer emitted
• Data-Code and Instance-Code of files written by image_meta_embed (byte-level changes from
  exiv2 0.19.2)