Decoding accuracy in Electroencephalographic (EEG) brain–computer interfaces is governed by how heterogeneous feature descriptors are fused: time-domain, frequency-domain, and time–frequency blocks are acquired with markedly different dimensionalities, scales, and correlation structures. Selecting and weighting these blocks, while the per-block regularization that controls the lower-level classifier fit is tuned in response, constitutes a bilevel optimization problem, with a leader fixing a fusion configuration and a follower returning regularized estimates. This paper develops HEF-Brain Storm Optimization (BSO), a bilevel optimizer whose swarm dynamics follow the BSO paradigm of idea clustering and idea generation, enhanced by a covariance-aware mutation operator defined on Symmetric Positive-Definite (SPD) matrices estimated from the heterogeneous feature blocks. Cluster-level covariance estimates are stabilized by shrinkage, safeguarded by eigenvalue clipping in [10⁻⁴, 10³], and coordinated with the lower level through warm-started follower solves governed by a 60/40 evaluation-budget split. A supermartingale-based convergence argument and an amortized complexity of O(Nd + N log k + d³/5) per iteration are supplied. The experimental program evaluates ten SMD bilevel instances over 30 independent runs against nine rivals, BLEAQ among them, and contrasts the learned fusion with the EEGNet, ShallowConvNet, and DeepConvNet decoders under leave-one-subject-out transfer on an eleven-subject motor-imagery cohort. HEF-BSO attains the smallest median optimality gap on seven of ten instances, with a suite-wide median of 4.4 × 10⁻⁶, and Pratt-adjusted Wilcoxon tests with Holm correction consolidate the margin. A hierarchical Bayesian model with subject-level partial pooling yields a posterior mean leave-one-subject-out balanced accuracy of 75.5%, against 74.1% for EEGNet, with posterior probabilities of superiority of 0.90 against EEGNet and above 0.995 against the remaining decoders. Perturbation and coupling-stress protocols hold degradations between two-fifths and one-half of the strongest rival's and bilevel infeasibility at 7.8% under a starved thirty-percent follower budget.