Improved PDF redaction stability so repeated processing produces consistent output.
Improved the LLM used to detect key-value pairs in PDF files.