inverted-pendulum family — per-instance multi-start synthesis (n_pend=20, k_inits=24, steps=100, θ0=0.25 rad) feasible (a tuned linear φ stabilises): 20/20 Among FEASIBLE pendulums: FAIL-to-stabilise rate across random inits: median 60% (mean 49%, max 75%) -> even on STABILISABLE plants, a random init lands in the DIVERGENT basin this often. cost spread among STABILISED inits: median 190% (max 495%) -> multiple local minima (>1 cost mode on 20/20 feasible pendulums). per-pendulum multi-start solve time: median 16.4s (24 inits × 100 steps). CONVEX CONTRAST — same multistart on LTI+sector loops (LMI-stabilisable): fail rate median 0% vs pendulum 60%; cost spread median 0% vs pendulum 190%. -> the LTI loop is reliably stabilised from random inits; the pendulum is not. INTERPRETATION: unlike LTI+sector (convex LMI), per-instance pendulum synthesis is non-convex — random single-shot synthesis fails often and multi-start is needed (slow/unreliable). An amortised generative prior supplies a good init / a distribution of stabilising controllers, which is exactly the value §6r's warm-start demonstrated.