Summary

This fixes a gateway restart race seen during hermes update when multiple hermes-gateway* systemd units are active. The change is not related to STT, Whisper, CUDA, or any local audio configuration.

Root Cause

get_running_pid() already determines when a PID file is stale, but cleanup for the current HERMES_HOME path delegated to remove_pid_file(). That helper intentionally refuses to delete a PID file owned by another process. For a stale PID record from a previous gateway process, this left gateway.pid behind and caused the next systemd start to fail repeatedly with PID file race lost to another gateway instance.

Separately, hermes update discovered active hermes-gateway* units in systemctl output order. If a profile gateway restarted before the default gateway, the default profile could hit stale PID state while another profile was already running.

Changes

• Unlink stale PID paths directly once stale cleanup has proven the recorded process is gone.
• Sort systemd gateway units so hermes-gateway restarts before profile units such as hermes-gateway-scout.
• Add a short pause between gateway service restarts to reduce same-update races.
• Add regression tests for stale PID cleanup and multi-profile restart ordering.

Validation

• PYTHONPATH=/tmp/hermes-agent-upstream-pr /home/tommaso/.hermes/hermes-agent/venv/bin/python -m pytest /tmp/hermes-agent-upstream-pr/tests/gateway/test_status.py /tmp/hermes-agent-upstream-pr/tests/hermes_cli/test_update_gateway_restart.py -q
• Result: 68 passed in 3.50s