- Removed set -e — agent no longer crashes when a remediation step fails
- SIGKILL/OOM → restart first, only rollback if exceeding max restarts/hour
- Added HEALTHY_CYCLES counter: after 10 consecutive healthy cycles (5 min),
reset restart tracking state (system considered stable)
- Rollback now handles the case where Docker restart policy already revived
the container (rm -f before re-creating)