CLASP: Constrained Lagrangian Adaptive Scheduling Policy (code and results)

Complete, reproducible code and results for the paper "Adaptive constrained reinforcement learning for energy-aware scheduling under changing load" (CLASP). CLASP is an energy-aware DVFS scheduler that holds an explicit SLA-violation constraint across load by driving a Lagrangian dual variable with a PID controller and conditioning the learned policy on the resulting constraint pressure. A single trained policy tracks a target violation rate across load and nonstationary shifts, can be retargeted to any SLA target at deployment with no retraining, generalizes to loads and bursty arrival patterns unseen in training, and extends to heterogeneous multi-class service objectives via a price vector. The repository contains the scheduling environment, the CLASP agent, all baselines and ablations, the full experiment suite including the revision experiments (dense weight sweep, controller comparison, PID-gain sensitivity, serve-late ablation, nonstationary tracking with baselines, zero-shot retargeting, generalization, and multi-class), a DQN function-approximation instantiation, the figure-generation code, and the numerical results. Implemented in pure NumPy and Matplotlib; no GPU or external data required.