We address fast rescheduling at a single station of an aircraft Final Assembly Line (FAL) under stochastic processing times and precedence constraints. Exact solvers (MILP/CP) provide high-quality plans but are too slow for online rescheduling after disruptions. We propose a Graph Neural Network (GNN) that, given a partial schedule and job-duration distributions, predicts the next job(s) to dispatch, the worker assignment, and a buffer time to reserve. The GNN is trained by imitating stable schedules produced by a CP-based Sample Average Approximation (SAA) with explicit buffers. On synthetic instances (30, 60, 120 jobs) the model reproduces CP decisions with high fidelity and yields competitive end-to-end schedules while enabling rapid inference suitable