Mycoplasma pneumoniae is a human pathogen recently proposed as a genome-reduced model for bacterial systems biology. Here, we study the response of its metabolic network to different forms of structural stress, including removal of individual and pairs of reactions and knockout of genes and clusters of co-expressed genes. Our results reveal a network architecture as robust as that of other model bacteria regarding multiple failures, although less robust against individual reaction inactivation. Interestingly, metabolite motifs associated to reactions can predict the propagation of inactivation cascades and damage amplification effects arising in double knockouts. We also detect a significant correlation between gene essentiality and damages produced by single gene knockouts, and find that genes controlling high-damage reactions tend to be expressed independently of each other, a functional switch mechanism that, simultaneously, acts as a genetic firewall to protect metabolism. Prediction of failure propagation is crucial for metabolic engineering or disease treatment.