Inconsistent results when defining the hessianstructure() method in the Problem interface

[lfaucheux]

A simple question. Given the difficulty of reproducing what I'm dealing with (I've tried in vain to come up with a concise example), I'm more in search of a hunch than a definitive answer.

Do you know why enrich scipy_interface.py at line 228 with

    [...]
    def hessianstructure(self):
        return np.nonzero(np.tril(np.ones((15, 15))))  # actually as done in ipopt_wrapper.pyx
    [...]  #                               ^^  my decision's dimensionality, hardwritten to save time. 

make my scipy.optimize.OptimizeResult-based result change from

...
 message: ...
 success: ...
  status: ...
     fun: 2.3124170059382245e-29
     nit: ...
    info: ...
...

to

...
 message: ...
 success: ...
  status: ...
     fun: 2.3000910542941462e-29
     nit: ...
    info: ...
...

The change is indeed epsilonian in this case, but it still shouldn't occur. More importantly, defining ~.hessianstructure prevents Ipopt from converging in subsequent resolutions. When digging in ipopt_wrapper.pyx, it appears at line 1094 that explicitly defining ~.hessianstructure as I did avoids fixing the "wrong stride" problem.

Any idea what's going on ?

---

OS : Windows 10
Ipopt version : 3.13.3
Python version : 3.11.6
cyipopt version : 1.6.0.dev0