Distribution is fine. The issue is likely one of:                                            
                                                                                
  1. Very few edits happening — only 20% of edits go to FireRed, and edits are a fraction of   
  total image requests                                                                         
  2. Queue check failing — from your GKE pods, the GET /api/queue call to Vercel might be      
  timing out (5s limit), returning False → silently skipping to CF                             
  3. No logs visible — the FIRERED_EDIT_ATTEMPT or FIRERED_QUEUE_CHECK_FAILED logs might be    
  there but you're not seeing them                                                             
                                                                     
  Can you grep your production logs for FIRERED? Even FIRERED_QUEUE would tell us if           
  should_use() is being reached at all. If there are zero FIRERED log lines, it means either no
   edits happened in those 20 min, or there's a deployment/import issue.