Behavior vs. Geometry — Drift Correlate Test

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Tests whether destroyed-binding's internal drift (Δ_role degraded under contradictory role definition, AC unaffected) has a behavioral correlate — does the model's actual generated output drift under contradiction, or does surface coherence conceal it entirely? Reuses destroyed-binding's exact coherent vs. contradictory conditions as the perturbation, and adds real text generation in place of hidden-state extraction. Primary hypothesis: Behavioral Fidelity Score is lower under contradiction, paired by role (Wilcoxon signed-rank).