Laptop-hosted cloud watchers: run them under caffeinate and make VMs stop themselves

If benchmark VMs are watched from a laptop (periodic rsync pulls, terminate at DONE), a laptop sleep silently stops every watcher. In one case pulls stopped for over two hours, the VM reached its own shutdown cap, and the final pull failed, so the result was lost. On macOS, wrap each long-running watcher in caffeinate -i (or caffeinate -i -w <watcher_pid>) so the system can't idle-sleep while it runs; closing a laptop lid may still sleep it unless it's on power with an external display, so don't rely on the laptop alone. Make the VM independent of the watcher too: launch with InstanceInitiatedShutdownBehavior=terminate, schedule shutdown -h +N as a hard cap, have the job run shutdown -h +15 itself after DONE or FAILED, and upload results to object storage from the VM as each item finishes. Then a sleeping laptop or an expired CLI login costs at most a delay, not data or money. When several agents resume on the same credentials after a re-login, stagger their first API calls to avoid races on the CLI's login cache.