fix(pdf): drop clipped and off-page text

[asionesjia]

Summary

Confirmed #79. The three cited one-page crops keep the next page's text in the content stream. Viewers and pdftotext -layout hide it with crop/clip. mdgate read the whole stream.

On 01030000000027.pdf that leaked 6.Conclusion, the following essay, and 49. On 01030000000028.pdf and 01030000000029.pdf each file mixed in the adjacent page, so the two markdown outputs overlapped.

CropBox equals MediaBox on these files. The leftover sits behind nested re W n clips (and, on the right-hand crop, at negative x). The extractor ignored both.

Changes

• Start each page with CropBox, else MediaBox, including boxes inherited from /Pages.
• Intersect that region with W / W* clip rects, saved and restored with q / Q.
• Skip text, strokes, rects, and annotations that do not overlap the current clip.
• Clip form XObject text to the form BBox.

Test plan

• bunx vitest run packages/pdf/test (images included after bun run build)
• bun run lint
• Pre-commit bun test (557 pass)
• Rechecked the issue's bench pages: 027 no longer emits Conclusion/49; 028 keeps 312 and drops 313/sections 5-6; 029 keeps 313/sections 5-6 and drops the previous page

---

Note

Medium Risk
Changes core content-stream interpretation and can drop previously extracted text if clip heuristics (axis-aligned path bounds) disagree with a viewer; scope is limited to the PDF package with targeted tests.

Overview
Fixes PDF-to-markdown leaking text that viewers hide via CropBox/MediaBox and content-stream clip paths (issue #79: adjacent-page content in cropped one-page exports).

The extractor now tracks a visible clip per page: CropBox (else MediaBox, including inheritance from /Pages), intersected with W / W* path bounding boxes and saved/restored on q / Q. Text runs, stroke lines, filled rects, form XObject content (intersected with BBox), and annotation items are omitted when they do not overlap the active clip; annotation appearances are skipped when /Rect is off-page, while in-page widget labels still render.

Adds packages/pdf/test/clip.test.ts covering MediaBox/CropBox, nested clips, form BBox, and annotation edge cases.

^{Reviewed by Cursor Bugbot for commit bb53d1e. 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	bb53d1e	Aug 27 2026, 07:20 AM

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Appearance clip mixes coordinate spaces
  • On-page appearances are kept by filtering on the annotation Rect in page space, and form streams are extracted without intersecting the form-space BBox with the page crop.

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

^{Reviewed by Cursor Bugbot for commit cc80126. Configure here.}

[cursor]

Stale comment

Left a non-blocking comment. Cursor Bugbot finished as skipped and reported 1 unresolved issue, so this is not approved. Human review is needed; no reviewers were assigned because no eligible non-author reviewer is available.

  

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

[cursor]

Approved. Cursor Bugbot passed on the current head with the prior finding resolved and no new issues, so no reviewers were assigned.

  

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