PR #1351 - Ephemeral Swarm Cloud Runners

Synlynk's swarm engine now has a small, explicit contract for ephemeral execution. A runner driver provisions isolated compute, streams telemetry, collects a receipt, and destroys the runner. The local driver makes this contract testable on a laptop; the Fly.io driver maps it to Machines v2 and starts a watchdog for every machine.

The manager keeps lifecycle records in SQLite, so synlynk swarm status is useful even after a CLI process exits. synlynk swarm dispatch fans out a batch, while synlynk swarm destroy --all is the emergency cleanup command. Progress is emitted as runner_progress relay events, giving the existing watch and relay surfaces a common stream without coupling them to a provider.

This is intentionally a foundation: Kubernetes and Hetzner drivers can be added without changing the CLI or the master-side lifecycle ledger.