Energy management in solar-based microgrids that also serve electric vehicles (EVs) in sub-Saharan Africa must cope with scarce measurement data, unreliable grid supply, volatile demand and tight battery and generator limits. Reinforcement learning (RL) can adapt to this uncertainty, but unconstrained RL policies give no assurance that operating limits are respected, and most published methods are validated on data-rich systems outside the region. This paper proposes a methodology, not a set of results, for a safe and data-efficient RL energy management system called SCOPE-EM (Shielded Constrained Policy Optimisation for EV-integrated Microgrids). It combines (i) probabilistic forecasting of photovoltaic (PV) output, load and EV demand with conformal calibration and transfer from data-rich sites; (ii) a constrained Markov decision process solved by a forecast-conditioned Lagrangian actor critic; (iii) a model-predictive shield that projects every action onto a chance-constrained feasible set; and (iv) an operator-facing explanation layer that records shield interventions. We specify the problem formulation, the data strategy, three archetypal case studies, the baselines and ablations, a pre-specified statistical protocol, and evaluation metrics covering cost, reliability, safety, EV service quality and equity. The hypotheses are stated so that the planned studies can refute them.