We present GreenCloud-RL, a formally defined scheduler for multi-cloud fleets that jointly optimizes carbon emissions, operational cost, and service-level agreement (SLA) risk. The scheduling problem is modeled as a constrained Markov decision process (CMDP) in which expected cost–carbon reward is maximized subject to aggregate SLA-risk and power-cap constraints. The framework integrates (i) a physics-informed DVFS power model, (ii) queueing-theoretic latency predictors, (iii) probabilistic short-term forecasts of regional grid carbon intensity and electricity price with uncertainty penalties, and (iv) a dual-variable actor–critic algorithm that provably enforces constraint satisfaction online. We analyze the algorithm’s convergence under bounded rewards and provide distributed complexity bounds for large-scale multi-cloud control. Empirical evaluation on production-like workloads shows that GreenCloud-RL reduces SLA-violation rate by 37.6% and power-cap violations by 53.8%, while lowering mean tardiness to 1.9 s and improving throughput by 4.2% compared with strong baselines. A region×regime analysis (load ∈ {low, medium, high}; grid ∈ {clean, neutral, dirty}) demonstrates consistent reductions in both cost and carbon emissions without degrading SLA satisfaction. Post-hoc calibration improves reliability (ECE 4.8%→1.7%, lower Brier score), enabling trustworthy on-time probability estimates for risk-aware admission and tenant-tier guarantees. Scalability tests show a p95 scheduling latency of ≈41 ms at 100k jobs, ≈0.25 ms inference per decision, modest CPU/GPU overhead, and ≈5 s controller failover. These results establish, both theoretically and empirically, that multi-objective optimization of carbon, cost, and reliability is achievable in production-scale cloud environments while maintaining contractual SLA guarantees.