i'd be betting my money on the overshoot of the reset pulse causing the problem. the overshoot, in turn, being caused by the 1k pulldown resistor charging up the 0.1uf series capacitor (by default, the 8U2 pins are open-circuit until configured).
a pulse that goes above the supply rail will do unpredictable things to the 328p, and the presence of your 4k7 pullup merely enables a set of conditions that direct the over-voltage pulse towards a lock-up condition.