Support for AI provenance metadata in PDF metadata streams

## Context AI-generated PDFs are increasingly common — reports, summaries, analyses generated by LLMs. But the PDF metadata (XMP, document info dict) carries no provenance about AI generation: which model, confidence level, sources used, or verification status. ## The Gap pypdf reads and writes PDF metadata, but there's no convention for AI provenance. As EU AI Act Article 50 (August 2, 2026) requires transparency metadata on AI content, a standardized way to store this in PDFs becomes important. ## Possible Approaches 1. **Custom XMP namespace** — store AI provenance in XMP metadata (pypdf already handles XMP) 2. **Document info dict** — add custom keys for `ai_generated`, `trust_score`, `model`, `sources` 3. **Adopt an existing format** — [AKF](https://akf.dev) defines a compact JSON schema (~15 tokens) for this that maps to PDF metadata streams ## Question Has the pypdf team considered AI provenance metadata? Is there a preferred approach for storing custom metadata beyond standard XMP fields? Would be happy to contribute if there's interest.