Block towers = skill; labels show skill / ceiling. Pace (tau) and ceiling (M) are illustrative model parameters, not measurements of any real child.
The science
Each learner follows a saturating practice curve S_i(t) = M_i (1 - e^(-t/tau_i)) (the solution of dS/dt = (M_i - S_i)/tau_i): a ceiling M_i and a time constant tau_i (smaller = quicker). With a budget T = sum t_i the total skill is largest when every learner who gets time has the same marginal gain dS_i/dt_i = (M_i/tau_i) e^(-t_i/tau_i) = lambda, i.e. t_i = max(0, tau_i ln(M_i/(tau_i lambda))); lambda is found by bisection so that sum t_i = T (reverse water-filling). Equal hours are generally not optimal, but neither is giving everything to the fastest learner.
Honest limits. Real learning curves are often closer to power laws, depend on prior knowledge and motivation, and differ by topic; an exponential with a hard ceiling is the simplest curve with diminishing returns. The numbers are illustrative and say nothing about any real student. Maximising a sum of skill is only one objective: fairness, wellbeing, interest and the value of a skill are not modelled, so the mathematical optimum is not a policy recommendation. The enrichment switch is a made-up trade (higher ceiling, slightly slower pace) to show how raising a ceiling changes the best split.
Checks. (1) Euler integration of the ODE against the closed form; (2) the water-filling optimum against 60 000 random splits (none may beat it); (3) equal marginal gains; (4) budget conservation. Scene: Blender-baked classroom (wood, plaster, brushed steel, tile) in Three.js.