OpenSCM v0.7.6

Fixed

• Agent (scmclient): stop duplicate/orphaned agents from piling up. Field hosts accumulated dozens of stray scmclient daemons over successive self-upgrades (an older restart path spawned a detached copy instead of replacing the process), and some grew to multi-GB RSS. The agent now takes a single-instance advisory lock (scmclient.lock, next to the config) at startup in daemon mode; a second instance logs and exits instead of running in parallel. The lock is O_CLOEXEC (Rust default), so it releases across the upgrade exec() and the re-exec'd binary re-acquires it cleanly — no self-deadlock. The run / --help / --version one-shots are unaffected.
• Agent memory: switched to a single-threaded Tokio runtime. The agent is a strictly sequential heartbeat loop but ran on the multi-threaded runtime, whose per-worker glibc malloc arenas never return freed memory — long-lived agents ballooned to multi-GB RSS / 20-GB VSZ. current_thread uses one arena, so memory plateaus low and flat. No functional change.
  • Operators: existing strays are detached from the service manager, so upgrading won't clear them. Clear once with systemctl stop scmclient && pkill -x scmclient && systemctl start scmclient (adjust the unit name).

---