Semantic projection scoring

Defines a bipolar 'feature line' as the vector difference between two antonym-pole word embeddings (e.g. 'small'→'big'), scores arbitrary words by their scalar projection onto that line, and validates the resulting scores against independently collected human graded judgments — an external, non-classifier correctness check on a single extracted direction.