perf: transcript Gemini fallback cleans up downloaded video on the event loop

[groupthinking]

Problem

TranscriptActionWorkflow._fallback_transcript_with_gemini cleans up the downloaded
video in a finally block that runs directly on the event loop.

Line	Call	Blocking work
transcript_action_workflow.py:852	video_path.exists()	stat() syscall
transcript_action_workflow.py:854	video_path.unlink()	unlink of a full downloaded video
transcript_action_workflow.py:857	temp_root.exists()	stat() syscall
transcript_action_workflow.py:858	shutil.rmtree(temp_root, ignore_errors=True)	recursive walk + unlink of the whole temp tree

_download_video_file requests
best[ext=mp4]/bestvideo[ext=mp4]+bestaudio[ext=m4a]/best with
merge_output_format: mp4, so the temp tree can hold the merged output plus
unmerged .fNNN fragments — potentially hundreds of MB across several files.
Each inode removal happens while the loop is stalled, so every other in-flight
request served by that worker is frozen for the duration.

The same file already does its download off-loop: _download_video_file builds a
synchronous _download() closure and returns await asyncio.to_thread(_download)
(line 1058). Only the cleanup half was left on the loop.

Reachability evidence

Reached from two live HTTP endpoints (router mounted at main.py:192):

POST /api/v1/transcript-action    router.py:456 -> run_transcript_action:462
POST /api/v1/videos/process       router.py:1489 -> _run_video_job:1515
  -> TranscriptActionWorkflow.run                   router.py:500 / router.py:1537
  -> TranscriptActionWorkflow.run                   transcript_action_workflow.py:99
  -> _extract_transcript                            :132 -> :307
  -> _fallback_transcript_with_gemini               :335 -> :708
  -> finally: exists()/unlink()/exists()/rmtree()   :852-858   <-- runs on the loop

Import-closure analysis from youtube_extension.main confirms the module is in the
production closure:
main -> backend.api.v1.router -> services.workflows.transcript_action_workflow.

Acceptance criteria

• No filesystem call in the finally block executes on the event-loop thread.
• Cleanup still removes the downloaded file and the temporary tree.
• Cleanup still runs to completion even when the caller is cancelled. The current
  synchronous code is uncancellable, so that property must not regress.
• Existing TestFallbackTranscriptWithGemini tests keep passing unchanged.
• Regression tests assert cleanup runs on a non-loop thread and that the loop stays
  responsive while cleanup is in flight.

Proposed fix

Move the four calls into a synchronous closure and await it via asyncio.to_thread,
mirroring _download_video_file. Preserve the predicates (exists() guards, the
OSError swallow, and ignore_errors=True) exactly — this change is about where the
work runs, not what it does.

[linear-code]

GRV-249

[github-actions]

👋 Thanks for opening this issue! It has been automatically labeled for triage.

[coderabbitai]

🔗 Related PRs

#314 - fix(pipeline): async kickoff uses videos/process [merged]
#496 - fix(youtube): complete youtube-transcript-api >=1.0 migration across remaining call sites [merged]
#648 - fix: address CodeRabbit review findings on merged perf pass (#642) [merged]
#1205 - perf(cloud-ai): run AWS Rekognition boto3 calls off the event loop [merged]
#1228 - perf: offload video result cache disk I/O to worker threads [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!