Efficient coordination of consumer-grade urban drone fleets requires joint optimization of task assignment, routing, and charging under tight onboard-energy, charging-infrastructure, and edge-computing constraints. This paper proposes a lightweight energy-aware multi-agent reinforcement learning (MARL) framework for consumer drone service networks. The problem is formulated as a constrained multi-agent sequential decision process that explicitly captures flight energy consumption, charging-station capacity limits, airspace traffic interactions, and long-term battery state-of-health degradation. Unlike a purely black-box MARL application, the proposed framework combines centralized training and decentralized edge execution with feasibility-preserving action masks, local communication features, and explicit battery-protection rules for safe reserve control, fast/slow charging selection, and charging-time scheduling. A compact actor network is executed on each drone using local observations, while a centralized critic is used only during offline training. A city-scale simulation environment over a 20×20 km urban area with 50 service nodes, 10 charging stations, and up to 200 drones is developed for evaluation. Compared with greedy, rule-based, MILP-approximation, single-agent RL, and representative MARL baselines, including MAPPO, QMIX, and COMA, the proposed method achieves the best overall performance, reaching a task completion rate of about 92%, reducing average energy consumption to about 1.7 kWh, lowering average service delay to about 2.0 min, and decreasing the charging congestion index to about 0.18. Sensitivity, complexity, and edge-feasibility analyses further show that the learned policy remains stable under reward-weight perturbations and can support real-time decentralized inference. These results demonstrate that the proposed framework provides a practical and scalable solution for sustainable consumer drone fleet management in future smart-city and consumer-technology environments.