Upper limit on R squared with two weak regressors

A response $y$ is regressed by OLS on $x_1$ alone and, separately, on $x_2$ alone (each with an intercept); both fits have $R^2=0.05$. What is the smallest number that is guaranteed to bound from above the $R^2$ of the joint regression of $y$ on $x_1$ and $x_2$ together?

$1$. Nothing prevents a perfect fit. For example, let $y$ be an exact quadratic in $x_1$ whose vertex is far from the origin and let $x_2=x_1^2$: each single regressor explains little, yet $y$ is an exact linear combination of $1$, $x_1$ and $x_2$, so the joint $R^2$ is $1$ (and individual $R^2$ values can be tuned to $0.05$).