Gibbs reweighting cannot leave the support

steering strength 1/α 0.56

base coverage width 0.08

achieved E[Q]–

in-support ceiling–

irreducible gap Δ–

Steel-blue = base π_base; teal = steered π^★ ∝ π_base·e^Q/α; amber dashes = value Q(a). The red shaded walls are outside the support (zero base mass) — the global optimum hides behind the right wall. Crank 1/α and the steered policy only ever climbs to the in-support ceiling; the gap Δ doesn't budge. Widen the base and Δ collapses.