A model trained on 50 features has coefficients that swing wildly between refits while its predictions stay stable. What does that pattern indicate?

D. Near-dependence among features, so many fits are almost equal

Multicollinearity: many weightings produce nearly the same predictions, so the fit is stable and the coefficients are arbitrary. The distinctive signature is exactly this split — reliable predictions, unreliable interpretation — and it is why regularisation exists to pick one solution from the near-degenerate family.