I am much more of a functionalist than Daniel Ross and Brian Collins, so I am much more sympathetic to unidirectionality, and the fact that there are counterexamples does not bother me.
It did bother Brian Joseph, who’s one of the big names against unidirectionality, and who also marked my thesis. He found it pretty good, despite the fact that it was written within the grammaticalisation framework.
Hypothesis is likely too strong a word, since it’s opened up to refutation through the counterexamples. But the tendency towards more grammatical, more reduced, more bounded, more obligatorified, is a thing, and it’s both quite useful in reconstruction (as Daniel concedes), and something to be explained.
Formalists in an online discussion on unidirectionality I read once dismissed unidirectionality in grammaticalisation as an epiphenomenon. The retort was: aren’t most interesting phenomena?