Over 350 million adults in the West African UEMOA region use Mobile Money as their primary financial instrument yet possess no formal credit history. Traditional credit scoring models systematically exclude this population — not because they are untrustworthy, but because the infrastructure to read their data does not exist. We present Fvundi, a hybrid ML-RAG credit intelligence system addressing this exclusion at three levels. First, a two-layer prediction-explanation architecture: XGBoost scores creditworthiness from Mobile Money transaction features, while a RAG pipeline (LlamaIndex + Cohere + Claude) generates faithful natural-language explanations grounded in SHAP feature attribution. Second, we identify and formalize explanation faithfulness failures — instances where the RAG explanation layer misrepresents the underlying ML model's decisions, introducing demographic inconsistencies for identical credit scores. We propose four quantitative metrics: Feature Coverage Rate (FCR), Feature Rank Correlation (FRC), Demographic Consistency Score (DCS), and Hallucination Rate (HR). Third, we introduce Constitutional Oversight: a three-layer governance framework combining a machine-readable Financial Constitution (human-authored ethical rules), Scalable Oversight (an AI Arbiter enforcing constitutional compliance at transaction speed), and Human-in-the-Loop control (triggered selectively for violations). This architecture satisfies BCEAO regulatory requirements while operating at the scale of millions of Mobile Money transactions per day. Our central argument: safe AI credit scoring in emerging economies requires institutional design that keeps humans in the loop at the governance level while delegating verification to machine-speed oversight. Fvundi is designed to extend the banking system, not replace it.