The 2026 FIFA World Cup was the first edition played by 48 teams across 104 matches. We analysed 142 official FIFA metrics per team per match, including line breaks, pressing events and receptions between the lines, none previously analysed at a World Cup. Win and loss performances were compared with Cliff’s delta under false discovery rate control; 86 of 125 screened metrics differed. Outcomes were modelled with five algorithms under repeated stratified cross-validation, using team-minus-opponent differentials. Explainability used SHapley Additive exPlanations (SHAP), and uncertainty used split-conformal analysis, a first World Cup combination. On-target attempts inside the penalty area showed the largest single effect (winners 4.4 versus losers 1.7; delta = 0.68). Win probability crossed 50% at 2.9 and 75% at 4.3 of these attempts. The regularised differential logistic model reached area under the curve 0.946 (accuracy 0.875), read descriptively because features exceed matches. Walking distance ranked first in averaged SHAP importance in the gradient boosting model; its strongest interactions joined total attempts on target (0.79) and on-target attempts inside the penalty area (0.71), the second-ranked feature. Teams dominating all five leading duels won 92% of decided matches; teams dominating none won 13%. Conformal coverage averaged 0.885 at the 90% target; 15% of matches remained uncertain. Finishing precision was the most consistent discriminator across stratifications (balanced-pairing delta = 0.53 inside the box, -0.54 conceded). Pressure-resistant possession showed strong overall and knockout-stage associations (delta = 0.56 and 0.65). Building these two capacities is the most defensible practical priority, pending prospective confirmation.