Classify whether a caption-matched crop is visually usable. This is intentionally conservative. A label/caption match proves identity, but no

Normalize a figure/table label to a canonical form for matching. 'Figure 3' / 'Fig. 3' / 'fig 3' / 'Figure. 3' all become 'fig 3'. 'Table 2'

Attach exact Obsidian wikilink matches to extracted reference candidates.

Return 'table' if the caption label starts with 'Table', else 'figure'.

Heuristic: a data row from a tabular layout, not part of a caption.

Keep only row cells in the same page column as a narrow caption. In two-column papers, unrelated left/right column tables often share the sam

Remove a leading YAML frontmatter block (---...---) if present.

Reject prose references such as "Table 8 summarizes ...".

Estimate the bounding box of the figure that lives ABOVE its caption. Strategy: 1. Collect all xref image rects and vector drawing rects on t

Walk away from the caption in ``direction`` ('up' or 'down') and collect logical rows that look like part of a tabular layout. Returns the li

Heuristic: a short textual row from a comparison/categorization table.

Gather bounding boxes of vector drawings on the page.

Gather the page-level bounding boxes of all embedded images.

Return caption anchors sorted top-to-bottom by their y0 coordinate. Each anchor contains the full multi-line caption bbox so that the downstr