chore: release 0.4.3

[maish]

Patch release: fix reprinted continuation-page headers appended as data rows on multi-page merge.

Fix

When a table's column header is reprinted at the top of each page — especially a multi-row (hierarchical) header — the repeated header rows survived the merge as bogus data rows, misaligning the stitched table for downstream consumers.

Injection now drops a body row when it is both flagged column_header by Docling and a tokenized match (Jaccard ≥ 0.6) for the reconstructed header block. Both signals are required:

• The flag alone is unreliable — Docling over-flags rowspan/continuation data rows as headers (e.g. repeated-header/rowspan-insurance-payout: 5 leading rows flagged, 3 are data). Trusting it alone deletes real data.
• The tokenized comparison is punctuation-agnostic, so per-cell OCR/VLM drift like (S$) vs ($$) is tolerated with no threshold tuning.

The merged DataFrame (lt.df) is unchanged; only the injected DoclingDocument is de-duplicated. A debug log reports each dropped row.

Tests

• 107 unit (1 new: drops reprinted header, keeps a column_header-mis-flagged data row).
• 54 integration (full snapshot lane). Three fixtures that encoded the old leaky output were re-baselined with an explicit, stricter injected_rows assertion: 15-page-druglist, covid-misc-labs ([3,4] table), retirement-portfolio. Every dropped row was verified to be a reprinted header (e.g. Sl No/Drug name/Category), never data.
• Verified against a real 3-page insurance-benefit table: 100 → 91 rows, body header pollution 9 → 0; header block, multi-level col_spans, and data integrity preserved.

🤖 Generated with Claude Code