v2.5.0 — Explainability (XAI) epic

Open the black box: two honest, model-agnostic techniques on the real Predict model, no heavy SHAP dependency.

• #105 Feature attribution — global permutation importance: shuffle each input, measure the accuracy drop, and see what the model truly relies on (and why permutation beats impurity importance).
• #106 Single-prediction explanation — occlusion from a typical-customer baseline turns one prediction into plain-language reasons and a per-feature push chart — a defensible per-decision explanation.

Both are labeled as approximations that ignore some feature interactions. Offline suite stays green at ~96% coverage.