fix(pdf): keep wrapped PDF table cells in their rows

[asionesjia]

Confirmed against the opendataloader-bench pages in #75. Cell assignment in packages/pdf/src/tables.ts dropped short row rules and treated wrapped lines as new rows, so headers leaked, multi-line cells split, and one grid became several GFM tables.

Changes

• Keep gridlines down to 12pt so short ruled rows still join the grid
• Add outer column edges from horizontal rules, and fall back to text rows when a grid has columns but only a top/bottom rule
• Fold wrapped continuation lines into the current row by column bands (not nearest column center)
• Allow one long text column so comparison grids with Y/N plus restriction paragraphs stay one table

Checks

• 01030000000081: TABLE 22 header and the wrapped Food Safety row stay in one 3-column table
• 01030000000045 / 46: org/party names and counts stay in columns
• 01030000000088–90: country, Y/N, and restriction text stay in one grid
• Corpus pdf/text.pdf snapshot unchanged
• bun test and bun run lint pass

fixes #75

---

Note

Medium Risk
Heuristic changes in table detection may mis-classify some PDF layouts (false tables or merged prose), though scope is limited to markdown output from packages/pdf.

Overview
Fixes PDF table extraction where wrapped text and partial ruling were split into extra rows or multiple GFM tables instead of one grid.

Ruled grids: lowers minimum rule length (MIN_RULE 24→12), extends column edges from horizontal rule extents, and when there are columns but fewer than two row bands, falls back to buildTextGrid (aligned-text folding) instead of rejecting the grid. Bounds use vertical rule extents where relevant.

Borderless / aligned tables: extends row runs with isContinuation (column alignment, caption skip, wrap span), then foldRows merges continuation lines into the current logical row via column bands (columnIndex / edges) and startsRow heuristics (keys, totals, lowercase wraps). Column inference is split into inferColumns + finishAlignedTable. acceptGrid and acceptBorderless are relaxed so one long column and larger tables (higher word/row limits) can still pass—e.g. Y/N plus restriction paragraphs stay one table.

Adds regression tests for short vertical rules, wrapped headers/cells, two-column wraps, and multi-column GATS-style grids.

^{Reviewed by Cursor Bugbot for commit 0b59ab7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	mdgate-demo	0b59ab7	Aug 26 2026, 07:48 AM

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

Bugbot Autofix prepared fixes for both issues found in the latest run.

• ✅ Fixed: Two-column wrap detection fails
  • inferColumns now divides matching two-column rows by full.length so wrap lines no longer dilute the 0.5 threshold.
• ✅ Fixed: Short wraps split table rows
  • startsRow now folds a single-column wrap into a cell that already has text, so short capitalized suffixes stay in the same row.

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit 188cffd. Configure here.}

[cursor]

Stale comment

Left a non-blocking comment and did not approve. Cursor Bugbot was present, then skipped, and its review still reports 2 unresolved medium-severity findings that need human attention. No reviewers were assigned because the only assignable user is the PR author.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[cursor]

Approved. Cursor Bugbot passed on this head with no unresolved findings that need human review. No reviewers were assigned.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}