Hi Alexey! I’m really glad to see you here.
The nasty signal I was illustrating is one where none of the samples exceed the ±1.0 rails, yet the “perfectly” reconstructed analog signal, x(t), has an overshoot that goes to infinity at t=-0.5 because the infinite sum of 1/n is unbounded and the polarity is never toggled in the sum (it’s all positive 1/n). If your upsampling filter is long enough, your +6.2 dB overshoot will be exceeded.