I have no opinion on the legal and ethical issues, I was commenting only on the classifier. I don't think your first question is appropriately described as "the real question" though, as one of the coolest results of machine learning is that there are techniques for combining multiple bad classifiers [0] in to one really good one.
Even more so, I think that is actually what is happening here. An officer is acting as a tier 1 filter, and then the drug dog is acting as a tier 2 filter, further reducing the number of false positives, which in a "legal and ethical" sense, we generally feel are worse than false negatives.