I've since improved it, and also discovered a new method of vector composition I have added as a first-class primitive:

debias_vector(axis, topic) removes the projection of axis onto topic: axis − topic * (dot(axis, topic) / dot(topic, topic))

That preserves the signal in axis while subtracting only the overlap with topic (not the whole topic). It’s strictly better than naive subtraction for “about X but not Y.”