Well, that was interesting... coffee break, so I went through to the Mac and gave it a spin.

If I use my "production" version of 1.3 (which is basically an early 1.3.5 candidate I froze for consistency) I get the sequence of events noted earlier - Rob's "working" sequence.

With either 1.4, or the 1.3.6 snapshot from git, I get the "failing" sequence.

So the change here is not in 1.4 per se, but somewhere in the 1.3.5 -> 1.3.6 progression.

That may also explain why Rob sees this effect; I'd guess he is using an older "stable"  version of 1.3.x for the bulk of his work?

Rob, can you give 1.3.6 a try to see if it also "fails" for you or not?