However, I completely agree that the developers don't want to be held down by any promises in this area. I would consider the exact number of times a variable is accessed to be subject to change.

Thanks. That's the important part. The rest is just quibbles about subtle shades of word meanings.

but other operators have similarly been changed as a feature (so the result is more in line with expectations), not as an optimization

The old behavior can't be what I call a "feature bug" according to ikegami who said:

The old behaviour is correct.

Yes, the improvement can be more than just an optimization. The new behavior can be considered better, but that behavior is not a documented nor documentable feature of join (because that would overly tie developers' hands). Which is what I mean by "not a feature bug".