OK, this was a tricky one. At least, I now understand the problem.
In the off-shell Higgs decay, the upper limit for the W* mass in the decay is mH - mW (the other W being on shell). This is 50 GeV for the given kinematics. In theory.
However, in practice the Higgs is resonant, but not forced on-shell. Hence the actual maximum allowed W* mass is the collider c.m. energy. Since Whizard generates squared masses, the ratio (mH-mW)/sqrts is squared.
In the phase-space channel which has the appropriate structure (Higgs resonance with off-shell decay), the decay events therefore populate only a tiny fraction of the initial range. VAMP can adapt to that, but it needs several iterations, and there is a chance that the interesting region is not found at all. Thus the Higgs-decay kinematics is sometimes discarded, and it gets worse if increased collider energy.
Even if I force Whizard to find it (by manually restricting the phs channels), there is a huge loss of efficiency if an event happens to fall just above the last bin boundary. This is an effect of the discretization. The result are fluctuations in the reweighting efficiency of order 100.
The lesson is that off-shell decays require a special mapping of the phase space of their decay products, which is different from the on-shell counterpart. I'll try to find an implementation.