`host_type` (`lxc`, `vm`, `proxmox`, `external`) is what the playbooks test to
treat containers and VMs differently.
## Playbooks
Always dry-run first: `-CD` (`--check --diff`).
| Playbook | Does |
|---|---|
| `utils.yml` | apt sources → `prism.viarezo.fr`, base packages, `qemu-guest-agent` on VMs. `--tags upgrade -l <hosts>`: full-upgrade one host at a time (never the Proxmox host); `-e utils_reboot=true` reboots VMs that need it, never LXCs. |
| `add_user.yml` | Accounts from `inventories/group_vars/all/users.yml`: create, keys (exact list), passwordless sudo, remove. Skips system accounts of the same name (Forgejo's `un-oeil`). |
| `deploy_kube.yml` | Longhorn prerequisites on workers, k3s install/upgrade/config one node at a time (init master → masters → workers), writes `kube/config`. Bump `k3s_version` in `group_vars/k3s.yml` to upgrade. |
## Secrets
`vault/password` (gitignored) holds the vault password. Nothing is vaulted yet;
the k3s join token is read from the first master. `kube/config` is gitignored.