feat: figure-level page-render extraction with caption-on-bottom table support

[KuangjuX]

Summary

This PR teaches extract_pdf_assets.py to render each Figure / Table from
the page pixmap with a caption-anchored bounding box, replacing (and
augmenting) the legacy xref-level extraction that often produced unusable
fragments such as 58×224 arrow icons.

The work is split into two commits so the table fix is reviewable on its own:

1. feat: add figure-level page-render extraction for complete figure cropping (df07e11)
   • Locates Figure / Table captions on each page via regex.
   • Computes a bbox that unions the embedded image rects and vector drawings
     between the previous caption boundary and the current caption.
   • Renders that bbox at 200 DPI and emits a new figure_assets list.
   • plan_figures.py matches plan items against figure_assets by
     normalized label and prefers them over xref candidates.
2. fix: correctly crop tables with caption-on-bottom and LaTeX cell fragmentation (740739a)
   • Tags each caption with kind: figure | table and stops merging
     multi-line captions when a numeric data row or oversized vertical gap is
     encountered, so the caption bbox no longer absorbs the first table row.
   • Adds row clustering (_cluster_lines_into_rows + _row_is_table_like)
     to fold the per-cell PyMuPDF lines that LaTeX tables emit (often 100+
     single-cell lines per table) into one logical row.
   • Adds a stricter _find_paragraph_blocks (≥200 chars, ≥3 lines, <40 %
     numeric-heavy lines) so table-shaped blocks are no longer mistaken for
     prose.
   • Splits bbox estimation into _estimate_figure_bbox_above_caption (legacy
     figure path, unchanged behaviour) and _estimate_table_bbox (probes
     both above and below the caption, picks the side with more table-like
     rows). Each estimator falls back to the other if its primary direction
     returns no usable bbox.
   • extract_figure_regions selects the estimator based on anchor kind and
     records kind on each emitted asset (additive change; no existing field
     of figure_assets is removed or renamed, so plan_figures.py /
     materialize_figure_asset.py continue to work unchanged).

Why a single PR?

The table fix relies on the figure-level scaffolding from the first commit
(_find_caption_blocks, _find_body_text_blocks, extract_figure_regions).
Splitting it would make the second commit unbuildable on top of upstream
main. The two commits are kept separate inside the PR so reviewers can
evaluate them independently.

Test plan

• Re-extracted the SkVM paper (arXiv 2604.03088): figure-level path
  produces 19 complete crops covering Figure 1–16 + Table 1, vs. 6
  fragmented xref outputs from the legacy path.
• Re-extracted the LoongTrain SC'24 paper (arXiv 2406.18485):
  Tables 2/3/4/5 now include the full column header, all data rows, and
  the complete "Table N." caption text. Previous output only captured
  the bottom data rows + the lower half of the caption.
• Verified Figures 2/4/6/8/9/10/11/12/13/14 on the same paper still
  extract correctly (no regression on figure-above-caption layout).
• plan_figures.py / materialize_figure_asset.py unchanged; existing
  fields of figure_assets preserved.

Notes for review

• figure_assets[*] gains a new optional "kind" field
  ("figure" | "table"). All existing fields are preserved verbatim.
• The legacy xref-level extraction is intentionally retained as a fallback
  so users who already depend on it are not affected.
• No new third-party dependencies; PyMuPDF (fitz) was already required
  and remains the only PDF backend.

[917Dhj]

Thanks a lot for working on this. I tested the PR on a set of papers, mostly CS/AI conference-style papers with many figures and tables.

My overall feeling is: I’m happy with the new extraction method, and I’d like to keep that part. It does improve the asset pool by adding captioned figure_assets, especially for papers where the old xref-only extraction misses most useful figures/tables.

However, I’d like the insertion and review logic to stay aligned with the original DeepPaperNote workflow.

The original design is intentionally placeholder-first:

• scripts extract evidence and candidate assets;
• figure/table placeholders are preserved first;
• extracted images still need to be reviewed;
• only images that pass the quality/semantic check should be inserted into the final note.

So I don’t think the PR should change the insertion decision from the script side. In particular, a label match between a planned figure/table and an extracted asset should not automatically set:

insert_mode: "figure_asset"

Some extracted assets are useful, but some are still duplicated, too tight, too loose, or not suitable for the final note. That is expected for an extraction step, but it means the output should remain a candidate pool rather than becoming an insertion decision.

Could you please keep the new figure/table extraction method, but adjust plan_figures.py so that it preserves the original placeholder-first behavior?

Concretely, I’d prefer:

• keep insert_mode: "placeholder" by default;
• attach matched figure_assets as candidates, e.g. figure_asset_candidate or candidate_assets;
• leave the final decision of whether to materialize an image to the existing review/model-side workflow.

That way we get the benefit of your improved extraction method, while keeping DeepPaperNote’s original insertion and quality-review logic intact.

A few edge cases I noticed during testing:

• duplicate labels, such as repeated Table 2 / Figure 3;
• inconsistent labels such as Fig. 1 and Figure 1;
• some table crops are still a bit tight or loose.

These are fine as candidate-generation issues, but they are exactly why I’d prefer not to automatically promote extracted assets into inserted note images.

So my requested change is mainly architectural rather than rejecting the extraction work: please keep the extraction improvement, but keep insertion/review decisions following the original placeholder-first workflow.

Thanks again — I think this is a useful improvement once that boundary is restored.