Thanks to @carlgleisner for the [detailed reproduction](https://github.com/carlgleisner/reactor_notification_worker_issue).
The problem was caused by an attempt to not have nested reactors indepdently publish their notifications separate to the parents but contained a logic flaw which caused the agent to not start, but only the first time you try and use a given reactor.
The fix involves _always_ starting a notification agent for each reactor, but nesting them. When a reactor completes it either publishes it's notifications to the parent reactor or to ash if there are no parent reactors.