> I didn't immediately spot the durability contract for volumes -- local NVMe like Fly, or sitting on some underlying system like EBS?

Persistent volumes use local NVMe that continuously syncs to object storage, so a similar architecture to Fly Sprites. It automatically recovers from host failures (with on-demand lazy restore), and you can manually migrate stable members to another region if needed.

I've positioned persistent volumes as an escape hatch for when you need an existing database app running in a VM, or when performance requirements outgrow the built-in replicated SQLite.