fix: promote implicit header rows into struct-tree tables

[abimaelmartell]

Summary

Some tagged PDFs mark only the data rows under <Table> / <TR> / <TD>, leaving the visible column headers above the grid untagged. Those headers then slip through to the heuristic fallback and reassemble into a spurious mini-table next to the real one (e.g. the Anthropic system card page 241 summary table: struct-tree covered 3 data rows but the header row was untagged).

After detecting a struct-tree table, scan unclaimed items within 60pt above its top. Items that start at one of the table's column left-edges are grouped by Y-row, coalesced per column across wrapped lines, and prepended as a single header row to the table's cells.

Guards

Protect against common false positives from page headers, captions, and stray text above the grid:

• ≥75% of columns must be filled (rejects stray page headers or event banners that only touch one or two columns)
• No cell may be a caption marker (Table 1, Appendix Table A2, Figure 2)
• No cell may be prose (starts lowercase, > 12 words) or suspiciously short (< 3 chars)
• No two filled cells may be identical (rejects banners that happen to align across columns)
• Non-alphanumeric fragments (a lone . or -) are skipped before merging so a stray period can't lead a row

Impact

• 1 pdf-evals snapshot updated — the Anthropic Mythos system card. Multiple "Summary of Claude's answers" tables now render with proper column headers instead of:
  1. an H1 mashing together the "Category" word with the title,
  2. a data table missing its header row,
  3. a stray 2-col mini-table below holding the wrapped header fragments.
• Composite score 0.679 (flat +0.001), char_accuracy 0.783 → 0.786, other metrics flat.
• Two new unit tests cover header merging and the false-positive rejection path.

Test plan

• cargo fmt + cargo clippy --release -- -D warnings clean.
• cargo test --release — 384 unit + 107 integration all pass.
• pdf-evals bench.py score — all metrics up or flat.

🤖 Generated with Claude Code