- Add logic to set master weight to 0 when first client connects (unless explicitly set via command line)
- This ensures proper load balancing between local and remote workers

Fix job ID mismatch in progress updates - use full job_id instead of job_id_int for database lookups

Implement progress tracking for jobs: add progress updates in worker, backend handlers, and database updates

Fix CLI debug argument to default to False instead of reading from database, preventing unwanted debug inheritance

Fix debug reset - only consider CLI args and env vars, ignore config files to prevent unwanted debug persistence

Fix debug reset logic - only consider CLI args and env vars as explicit, ignore config file settings

Fix debug reset logic - ensure debug and debug_web values are actually saved to database after reset

Apply same debug reset logic to --debug-web - reset to false at startup unless explicitly set via CLI, config file, or environment variable

Add real-time GPU usage column to cluster nodes page - displays live VRAM usage and GPU utilization with visual progress bars between GPU Memory and Workers columns

Fix race condition where jobs are sent before workers register - backend now returns error for unavailable workers, cluster master re-queues jobs with delay

Add real-time GPU usage monitoring to cluster nodes - clients send GPU stats every 5 seconds, master stores and serves via API, web UI shows usage bars and utilization

Fix GPU memory and utilization stats to show actual values using pynvml instead of PyTorch-only stats

Fix backend to always return response for job forwarding - success/error messages ensure proper job assignment feedback