Phase 19-20 branch-isolation principle preserved — portfolio-level constraints disabled at branch level, only deadzone-gated per-symbol anti-pattern regulators (action_l2, action_diversity, direction_inbalance) remain. Distributional CVaR penalty and q_crossing identical to v2.4.4.
No CSC knobs. csc_momentum / csc_eps / csc_apply_* are absent — there is no CrossSampleCenter to configure.
Role as control. v2.4.4-mini closes the 2×2 grid and gives the experimental matrix a clean comparison framework:
full d_model 128 : v2.4.4 v2.4.18 mini d_model 64 : v2.4.4-mini v2.4.18-mini
Horizontal: isolates the CSC effect at each capacity scale. Vertical: isolates the capacity effect at each architecture. Diagonal: exposes capacity × CSC interaction effects.
Hypothesis-testing reads. If CSC improves only at full scale, it is a size-dependent intervention. If the CSC effect is consistent under compression, the mechanism is architecturally robust. If v2.4.4-mini already matches full v2.4.4 on diagnostics, the full architecture is operating with significant capacity reserve.
Deployment status. v2.4.4-mini is a methodological control, not a deployment target. Run in parallel with v2.4.18-mini to fill the two lower cells of the matrix; combined with archived full-model results, this produces the complete 2×2 dataset for attributing architectural effects.