Can anyone suggest forcing monotonicity in noisy data decreases the accuracy of XGBOOST?
XGBoost constructs a model by optimization. Monotonicity enforces a constraint on that model, and creates a constrained optimization problem. It's always true that constrained optimization problems have optima which are are at best as good as the unconstrained problem; if the constraint is active, then the optima must be worse than in the unconstrained problem.
Stated another way, monotonic models enforce that increasing (decreasing) the value of a feature can only increase (decrease) the value of the response. Sometimes we want to do this because we know it's physically impossible for some process to increase-then-decrease-then-increase, or we have other relevant knowledge. But if oscillation truly a part of process we're modeling, then monotonicity constraints will obviously make the model worse.