Is CE search path properly decoded for latent space search?

[pat-alt]

Working in an old version of the code base (revisiting an old paper) and noticed an inconsistency with TaijaPlotting for CEs trained with latent-space search (path shown is the path in the latent space).