Fix: segfault when uploading multiple PDF attachments concurrently

[lauty95]

Problem

Uploading two PDF attachments to a chat close together can crash the whole bench start process. Not recoverable — needs a full restart.

Screen.Recording.2026-08-07.at.12.37.21.PM.mov

Error

Fatal Python error: Segmentation fault
...
File "flow/knowledge/extract.py", line 212 in _render_page_png
File "flow/knowledge/extract.py", line 178 in _extract_pdf_page
File "flow/flow/doctype/flow_session_attachment/flow_session_attachment.py", line 56 in extract_attachment

The frontend fires one upload_file request per attached file, so two PDFs upload as two concurrent requests, each extracting text synchronously on its own thread

pypdfium2 wraps PDFium. Two threads calling into it at once corrupts its state and segfaults the process.

Fix

Serializes all PDF processing in flow/knowledge/extract.py behind a single process-wide lock, so two PDFs uploaded together now get processed one after the other instead of racing.

Screen.Recording.2026-08-07.at.12.38.39.PM.mov

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge.

All PDF rendering paths are covered by the lock, and the regression test exercises concurrent access.

_{Reviews (1): Last reviewed commit: "fix: serialize PDF extraction to prevent..." | Re-trigger Greptile}