Agent-drafted interpretation, pending human review

This sweep behaves like a clean reference tradeoff rather than a brittle failure case. In the discrete model, mode recovery is strongest at lower beta and degrades gradually as KL pressure grows. In the continuous model, linear decodability of the true modes remains almost perfect across most of the sweep while latent geometry compresses toward the prior.

The main comparative lesson is that stronger regularization pushes both models toward prior matching, but it harms them differently. The discrete model loses some mode-recovery quality early, while the continuous model preserves linearly decodable structure much longer and primarily shows up as a shrinking latent radius before probe quality finally drops at the largest beta.