Headway switching, rather than feedforward loss alone, destabilises cooperative platoons under V2X dropout

Cooperative Adaptive Cruise Control (CACC) uses vehicle-to-everything (V2X) communication to sustain short, string-stable headways, but real wireless links drop packets, eroding this guarantee. We show that at a representative moderate-loss point, discontinuous headway switching on packet loss, not feedforward loss itself, is the dominant destabilising mechanism: a canonical plant-inverting design remains string stable losing feedforward at fixed headway, but not when switching headway on loss. Continuous, non-learned scheduling recovers stability at moderate but not severe loss. We train a dropout-aware multi-agent reinforcement-learning (TD3) controller with a learned V2X reliability gate; four of five training seeds converge to a policy string stable at $$30\%$$ and $$60\%$$ loss, outperforming the strongest classical scheduler under V2X-only information. The comparison is not information-matched: the learned policy also observes an idealised, always-exact onboard predecessor acceleration no classical controller here uses, and a non-learned controller given that signal alone is competitive with it, limiting attribution to learning. Frequency sensitivity, severe-burst robustness, and training reliability are all seed-dependent. These results identify switching, not information loss, as a primary vulnerability of communication-dependent platooning, and give a candid account of how much a learning-based remedy owes to learning itself.