Rerunning the cluster role recreates node containers with Compose v2.38.2 on Docker Engine 28 #5

Rerunning the cluster role recreates every k3s node container when the host runs Docker Compose v2.38.2 on Docker Engine 28.0.4 (GitHub's ubuntu-latest). The role's "Start k3s via docker compose" task reports changed on the second run and all three node containers get new start times, which restarts k3s on every host each time the playbook runs.

Seen in the mesh test's hosted and in_cluster scenarios on ubuntu-latest, before its workflow started installing Compose itself. Not seen with:

• Compose v5.5.1 on the same runner (Engine 28.0.4): the mesh test's no-restart check passes
• Compose v2.38.2 on Docker Desktop: also passes

So it's that Compose version together with that engine, and I haven't found which part of the service configuration it thinks changed. The CI workflows now pin Compose, so the tests no longer catch this for anyone running the role with an older one.

Either the role should refuse Compose versions older than one known to be stable across reruns, or the cause should be found and fixed. Repro: run tests/mesh/run.sh hosted on ubuntu-latest with its preinstalled Compose.