swm: one-command GPU rental with workspace persistence across providers

swm solves the perennial pain of renting cloud GPUs: each session starts with 45 minutes of reinstalling custom nodes, models, and configs. The tool unifies GPU discovery across providers like RunPod, Vast.ai, Lambda, and seven others with a single `swm gpus` command that sorts by price. Once you pick a GPU, `swm pod create` spins up the instance, and `swm setup install comfyui` (or Ollama, OpenWebUI, vLLM, SwarmUI, Axolotl) installs the framework in seconds.

The killer feature is workspace sync: your entire setup (models, custom nodes, outputs, configs) lives in S3-compatible object storage (e.g., Backblaze B2). When done, `swm pod down` pushes everything and kills the instance. Next session on any provider, a simple pull restores the exact state. A lifecycle guard watches GPU utilization and auto-saves/terminates after 30 idle minutes (configurable). Background auto-sync pushes changes every 60 seconds, and tar mode bundles many small files into one S3 object. The tool works with AI agents (Cursor, Claude Code, Codex, Windsurf) and is free, open-source, and Apache 2.0 licensed.