Docling: scanned PDFs fail as invalid_provider_output, and PDFs with pictures lose page units

Summary

doclingSourceDocument uses Docbank's page evidence only when it is complete. Docbank marks Docling evidence partial whenever the result contains pictures, so msgvault falls back to the provider Markdown. The Markdown leaves out text inside pictures, which is where Docling puts OCR text from a scanned page. Scans end terminal as invalid_provider_output; other PDFs with any picture index as one document unit instead of pages.

In my archive: 51 scans fail this way, and 4,620 of 4,805 indexed PDFs have no page units.

Environment

msgvault d224f069 (function unchanged on d8534b24), docbank v0.14.1-0.20261001015536-a212ec3d2e4a, local docling-serve-cpu:latest with default options, darwin/arm64.

Reproduce

Index any image-only scanned page. Sent directly to Docling Serve, a scanned receipt returns <!-- image --> as md_content and 28 OCR'd lines in json_content.texts, all with page provenance. Through msgvault the same file ends terminal.

What I expected

Use partial page evidence when its only omissions are unmapped structured fields, keeping page locators. Fall back to Markdown only when there is no usable page evidence.