Implement OCR pipeline from AutomatedFOIA for scanned document intake

[stuckvgn]

Context

The current ingester in src/documents/ingester.py has an OCR fallback path that calls
_extract_text_ocr(), but this requires pdf2image + pytesseract as optional extras
and system packages (tesseract-ocr, poppler-utils).

Work Needed

• Document exact system dependency installation steps for Linux, macOS, and Windows
• Test the OCR path against representative scanned USDA APHIS inspection PDFs
• Verify the in-memory decryption → OCR pipeline (BytesIO, never writes temp files)
• Add per-page OCR fallback: only fall back on pages where native text is empty, not the whole doc
• Add quality check: if OCR yields < N chars on a page, flag for manual review
• Wire up to src/api/server.py document ingest endpoint
• Add integration test with a sample scanned PDF (use synthetic/non-sensitive test fixture)

Security Note

OCR must process in RAM only (BytesIO). No unencrypted temp files — this is a hard requirement
for ag-gag jurisdictions where device seizure is a real threat.

References

• AutomatedFOIA/backroom.py — reference implementation (has Windows-specific paths to clean up)
• src/documents/ingester.py — current implementation
• docs/security.md §Encrypted Storage