I agree that of course one event handler cannot interrupt another.
This is the difference between an event loop and preemptive multi-tasking or true multi-processing. I was thinking of the latter as a comparison.
I don't understand what you mean by "undefined timing" compared to the sentence following it.
I consider a race condition to be the situation when events in one system occur in a non-deterministic order — and the exact order in which the events occur matters, which is the bug. Race conditions are solved by either adding additional synchronization to the system to ensure the events occur in a useful order or ensuring that the system operates correctly regardless of the event ordering.
The model you describe of "set everything up before the event loop starts" doesn't work in network servers for example;
You are correct, but that is the model I am currently familiar with (I have not yet had need to write any low-level network servers) and explains the blind spot I had that you have helpfully pointed out.
And your solution is better; I had not noticed the use of a quasi-global lexical variable in the question until now. I agree that getting rid of $TX is an important step in improving this code.