Properly learning decision trees in polynomial time?

Let $n,s\in\mathbb{N}$ and let $f:\{0,1\}^n\to\{0,1\}$ be an unknown Boolean function promised to be representable by a rooted Boolean decision tree with at most $s$ leaves (i.e., each internal node queries one input bit, and each leaf is labeled in $\{0,1\}$). Let $\mathcal{U}$ denote the uniform distribution over $\{0,1\}^n$. A randomized proper learner is given $(n,s,\epsilon,\delta)$ and oracle access to membership queries for $f$ (on any adaptively chosen $x\in\{0,1\}^n$, the oracle returns $f(x)$); since $\mathcal{U}$ is known, the learner may also generate independent samples $x\sim\mathcal{U}$. The learner must output an explicit decision tree hypothesis $h:\{0,1\}^n\to\{0,1\}$ such that, with probability at least $1-\delta$ over its internal randomness, the uniform error is at most $\epsilon$: