The Root Shapes the Fruit: On the Persistence of Gender-Exclusive Harms in Aligned Language Models