bug: PDF: the document structure of long mixed native-and-scanned PDFs holds only paragraph nodes, no headings, tables or page breaks

Description

include_document_structure = true on three long PDFs with mixed native and scanned pages returns a document tree that holds nothing but paragraph nodes:

PDF	tables in tables	structure nodes	node types
A, long, mixed	14	8,146	paragraph only
B, long, mixed	490	34,683	paragraph only
C, short, native	60	1,054	paragraph only

The same build returns heading nodes with levels, group, table, page_break and bounding boxes on a synthetic reportlab PDF with 24 pt and 16 pt Helvetica headings. On the real files there is no heading, no table (although hundreds of tables were extracted into tables), no page_break, and no list.

Expected: the tree carries the tables the extractor already found, page breaks between pages, and headings where the font metrics or the tag tree support them. A tree of paragraphs alone does not let a caller audit structure.

Steps to reproduce

1. Extract a long PDF with mixed native and scanned pages with {"pages":{"extract_pages":true},"include_document_structure":true} and the default OCR config.
2. Count document.nodes by content.node_type.
3. Compare counts.tables with the number of table nodes.

Relevant files and configuration

xberg v1.2.4, release build, 2026-09-18. Any long public-sector PDF with many tables should reproduce the table gap. extraction_method was mixed for the two large files and native for the short one.