fix(docling): drop off-by-one page-number shift in DoclingBackend

[cursor]

Bug and impact

DoclingBackend produced wrong page numbers for every block in every parsed document. A block on page 1 reported page=2, the last page reported one past the end of the document, etc.

Downstream effects across the project's stated use cases:

• RAG citations — every "see page N" reference is shifted by one, sending users to the wrong (or nonexistent) page.
• bigos parse --format=json — the JSON output that ingestion pipelines persist contains incorrect page fields.
• Search / page-anchored UIs built on Document.blocks highlight the wrong region.

This is silent data corruption: nothing crashes, no error is logged — every block just lies about its provenance.

Root cause

In src/bigos/backends/docling.py, _page_1_indexed_from_item(...) returned n + 1, on the assumption that docling's ProvenanceItem.page_no was 0-indexed.

In docling 2.92 (and the installed docling-core 2.74) prov.page_no is already 1-indexed: it is used as-is to index DoclingDocument.pages (e.g. doc.pages[prov.page_no]), and a single-page document's pages dict is keyed {1: ...}. Adding +1 on top therefore shifts every block.

Reproducer (single-page PDF fixture in tests/fixtures/simple_text.pdf):

src = Source(uri=pdf.as_uri(), mime_type="application/pdf", sha256=sha256_file(pdf))
doc = await DoclingBackend().run(src)
print({b.page for b in doc.blocks if b.page is not None})
# before fix: {2}
# after fix:  {1}

The existing tests only asserted block-kind counts, not page numbers, so the bug slipped through the initial docling integration review.

Fix and validation

• Pass the docling page number through unchanged after int(...) coercion. Treat < 1 (and missing/bad values) as unknown.
• Document the 1-indexed convention in the helper's docstring so future edits do not re-introduce the shift.

Validation:

• New focused unit test test_page_helper_passes_through_1_indexed covers the in-range, missing-prov, and bad-value branches without requiring docling at runtime.
• New slow regression test_parse_simple_text_page_number_is_one runs the docling backend on the existing single-page fixture and asserts every block reports page == 1.
• Full test suite (uv run pytest) — 67 passed, no regressions.

The fix is minimal (1 file, ~10 lines) and does not touch any other behavior.