In the registered staged v6/v7 gate-only ablation (MiniLM; $m = 100$, $k = 10$), with diffusion and role-aware quotas held fixed and only the contrast term of the adaptive controller toggled, LectureBank-Full moves from level-only R@$10 = 86.4$ $[82.4, 90.1]$ to level+contrast $87.4$ $[83.4, 91.1]$. The valid comparison is the paired-bootstrap delta between the two rows: $+1.05$ R@$10$ percentage points, 95% CI $[-1.52, +3.73]$, $p = 0.448$, $n = 143$. On MOOC-CS, level-only and level+contrast are tied at $23.1$ $[17.0, 29.6]$. Because the CI straddles zero and $p = 0.448$, the $+1.0$-point difference is not statistically distinguishable from zero; the ablation supports a modest, dataset-specific contrast-gating effect rather than a durable headline gain, and diffusion and role-aware quotas remain the main signal.