Abstract: The study proposes an optimization framework for heterogeneous kitting policies based on PI- containers to address buffer accumulation, production interruptions, and material returns caused by stochastic material loss during final assembly. The assembly and replenishment processes are first modeled as a Markov chain with absorbing states to characterize how material loss drives buffer evolution and triggers returns. Based on this model, a simulation–optimization framework is developed to compute return-triggering probabilities under different initial states and thereby identify the optimal kit composition. Several kitting policies are further designed to compare the effects of kit-type diversity and loss parameters on system performance. Numerical experiments based on a cigarette assembly line show that, compared with traditional fixed kitting policies, the heterogeneous kitting policy can reduce material-return probability by up to 10.43% and extend continuous assembly time. Within the tested parameter range, three to four kit types are sufficient to cover most effective needs, and adding more types yields limited marginal benefits. The study provides a new modeling framework and practical reference for kitting-strategy design under uncertainty.