e2e-cluster.yaml

Triggers: manual dispatch, pull requests (opened/synchronize/reopened, excluding doc-only path changes).

Runs the “staging” half of the multi-machine end-to-end suite (test/e2e/cluster): it provisions a bootloose cluster of Docker machines, uploads the cargoship package, and renders the engine config on every OS family - then stops before actually starting the engine anywhere.

This is intentionally split out from e2e.yaml into its own workflow so its trigger, concurrency group, and time budget are independent of the checks required on every push. It calls the cargoship packages directly (distro.Create, action.NewPrepare/action.NewApply) rather than a built binary, so it needs no build step.

The other half - actually installing k3s and bringing the cluster fully up - is not run in CI. It isn’t reliable enough on hosted runners to be a required check (even after switching from a three-controller RKE2/etcd topology to single-controller k3s to remove quorum timeouts). That half still exists and runs locally via mage test:endToEndCluster. See docs/dev/e2e-phase-tests.md and docs/agent/choice-e2e-stage-split.md.

On failure, node diagnostics (disk, memory, container logs) are collected and uploaded as an artifact, and leftover bootloose containers are always cleaned up.