The snapshot/restore detail is the most interesting part to me. If artifacts really suspend to storage after inactivity and restore on next access, that's the same bet Fly.io made with Firecracker machines, and it mostly worked once restore times got under a few hundred ms. I'd love to see the author measure actual restore latency across artifact sizes, because that number decides whether this pattern can replace always-on preview environments. There's also a cost angle worth digging into: per-session ephemeral VMs are cheap to suspend, but storage and restore I/O for big node_modules trees add up fast at scale.