The present paper is motivated by a partnership with Renault, who must routinely solve IRP instances of unprecedented continental scale and complexity as part of their backward logistics.Indeed, they receive car parts from suppliers at their plants in packaging, and reuse the latter, which implies the need for backward packaging logistics. The goal of our partnership is to redesign their IRP algorithm. We embed it into a multi-stage optimization framework. Indeed, only the short-term decisions must be taken on each day in practice, based on the current information of the future. We design a large neighborhood search (LNS) to solve the fixed-horizon deterministic multi-attribute continuous-time IRP. To this end, we derive a new flow relaxation, we generalize dozens of neighborhoods from the routing literature to the continuous-time IRP, we define two new perturbations based on new MILP formulations, and a new large neighborhood deriving a localized multi-attribute generalization of a matheuristic. We show we scale to the European instances through extensive numerical experiments. From this LNS, we define a policy to address the dynamic IRP, leveraging statistical models. This policy is currently being industrialized at Renault. We implement a simulator to evaluate our policy in a realistic framework, and to compare it with the current algorithm used in production at Renault. We highlight the potential gains: thousands of tons of CO2 and millions of euros per year.