ok, I think I know what it is: the kernel _activity_stream created here takes a finite time to subscribe, but in the test, the execution is processed right away, before that subscription propagates.

The kernel connection waits for its iopub channel to be fully connected via the nudge mechanism to resolve this same issue, but since it's not applied to the _activity_stream socket, there's always a chance that one won't be

So what happens:

1. kernel start requested
2. kernel execution request sent
3. kernel starts
4. kernel processes execution
5. replies and iopub messages finish
6. _activity_stream subscription resolves, but no iopub messages are forthcoming anymore
7. the test waits for _activity_stream to receive a message and update .execution_state

If we include checking kernel.execution_state in the nudge, I think that might fix it.

@krassowski I think this may have fixed Windows CI (at least one pass), by making sure the execution_state is set during the nudge.