How small (in modulus) can a polynomial get?

Yes. Proved by Chebyshev. The extreme case ($\max = 2^{1-n}$) is given by $2^{n-1}\cdot f(x)$ being a Chebyshev polynomial of first kind. https://en.wikipedia.org/wiki/Chebyshev_polynomials

You are essentially asking to approximate $x^n$ on $[-1,1]$ by a polynomial of degree less than $n$. The Remez algorithm will find the best possible aproximation. In Maple you can enter

numapprox[minimax](x^n,x=-1..1,[n-1,0]);

If you take $n=10$ or so and work to 100 digits you will quickly convince yourself that the maximum error is precisely $2^{1-n}$. It also looks as though the optimal approximation $f(x)$ has $2^{1-n}f(x)\in\mathbb{Z}[x]$. One could probably find a formula for the coefficients experimentally with the help of OEIS and then work back to a proper proof. But sadly I need to go and do less interesting things now.