During self-selected overground walking, we examined participant-level surface electromyography (sEMG) descriptors associated with Gross Motor Function Classification System (GMFCS) levels I–III in 38 ambulatory children with spastic cerebral palsy. Eight-muscle sEMG was summarized using a design-locked pool of 16 peak-normalized envelope and exploratory synergy features. Higher GMFCS level was associated most consistently with longer activation duty cycle, stronger distal and proximal peak-normalized co-activation, and greater inter-muscle spatial entropy. Directions were retained across sensitivity analyses. For secondary within-cohort discrimination, the 16-feature pool yielded a quadratic weighted kappa (QWK) of $0.837\pm 0.046$ across 50 repeated participant-level nested cross-validation evaluations. Repeating correlation reduction and four-feature selection inside every outer-training fold yielded QWK $0.849\pm 0.029$. Both estimates reuse the 38 children and describe resampling performance, not independent-population generalizability. The core-four specification was post hoc. In an open supplementary workbook containing 195 unique participant IDs, directional concordance was strongest for activation duty cycle and distal co-activation, but direct transfer of the internally fitted 16-feature and core-four models was poor (QWK 0.210 and 0.077). Without healthy controls or measured walking speed and stride length, these cross-sectional results identify descriptors associated with ambulatory GMFCS level, not cerebral-palsy-specific biomarkers or evidence of clinical predictive utility.