Hey Chris,
I checked with Itzik and he agrees with my analysis, but the reason behind his statement in the article is that he classifies SELECT...INTO under FULL recovery as not being truly minimally-logged since writing the changed pages is not deferred, and BCM bits are not set.
So, the explanation I offered stands, it's just a question of interpretation. For me, the main point is that SELECT...INTO is very much faster than the equivalent INSERT due to the logging differences.
I hope that covers everything - including why your tests and Itzik's article are both 'correct'.
Paul