C++ __enzyme_autodiff modifying arguements in unexpected ways

I have the following snippet that has been working consistently:

```C++
using ModelFunc = std::tuple (*)(double, double, double, double, const alloys::Alloy&);

template
inline double wrapper(double V, double R, double dT, double C0, const alloys::Alloy& A)
{
std::tuple f{MODELFUNC(V, R, dT, C0, A)};
return std::get(f);
}
struct Jacobian{double df1dV{}; double df1dR{}; double df2dV{}; double df2dR{};};

template
inline Jacobian calculateGrads(double V, double R, double dT, double C0, const alloys::Alloy& A)
{
Jacobian J{};
J.df1dV = __enzyme_autodiff(
(void*)wrapper,
enzyme_out, V, enzyme_const, R, enzyme_const, dT, enzyme_const, C0, enzyme_const, &A
);
J.df1dR = __enzyme_autodiff(
(void*)wrapper,
enzyme_const, V, enzyme_out, R, enzyme_const, dT, enzyme_const, C0, enzyme_const, &A
);
J.df2dV = __enzyme_autodiff(
(void*)wrapper,
enzyme_out, V, enzyme_const, R, enzyme_const, dT, enzyme_const, C0, enzyme_const, &A
);
J.df2dR = __enzyme_autodiff(
(void*)wrapper,
enzyme_const, V, enzyme_out, R, enzyme_const, dT, enzyme_const, C0, enzyme_const, &A
);
return J;
}
```

However my latest ModelFunc, which calls member functions of one of its arguements (A), has been returning incorrect gradients.

On debugging, i found that each of the 4 __enzyme_autodiff calls here are somehow modifying one of the member variables of A.

The ModelFunc and wrapper functions cannot modify A as they take A by const reference, and all of A's member variables and member functions are const.

However I could never get the snippet above to run without passing A by adress to __enzyme_autodiff.

Is there a way to pass class objects to __enzyme_autodiff by copy / const reference / pointer to const or otherwise prevent __enzyme_autodiff from modifying objects passed to it?

Could this be a bug aswell? If so, I can try reducing this to a simpler, reproducable example.

Start with the provided C++ snippet, especially the __enzyme_autodiff calls, wrapper, and ModelFunc member-function calls. Reduce it to a reproducible example that records A before and after differentiation, then determine whether the mutation is expected or indicates a bug; done means the behavior and required usage or fix are clearly established.

Written by the indexing model from the issue text.