teknium1 closed #104740 through #104887, merge 50cd1190ef2c. The restart client used to die at 15 seconds. _systemd_restart_timeout now reads TimeoutStopUSec and TimeoutStartUSec, adds 15s slack, and substitutes 90s per unknown or infinite phase.

I ran the committed harness evals/update_unit_client_budget.py on 2237be355906, not hermes update. Ordinary path: ExecStop=/bin/sleep 16, TimeoutStopSec=45, 16.137s, exit 0, PID 489559 to 489881, active. Catch-up: sleep 31, 31.153s, failed_units empty. Missing-unit commands still exited non-zero and wrote stderr on both runs. Cleanup stopped both units inactive. That tree’s tests/hermes_cli/test_update_unit_client_budget.py passed 13/13 in 0.28s this slot. origin/main 966637323e still has the same update_cmd_fleet.py.