LB cells and slices do not check for valid property-setters

[reinaual]

Our LB nodes and slices allow the user to set any value without throwing an error. The getters do throw.

import espressomd
import espressomd.lb


system = espressomd.System(box_l=[3, 3, 3])
system.cell_system.skin = 0.1

lbf = espressomd.lb.LBFluidWalberla(kinematic_viscosity=1.0, density=1.0, agrid=1.0, tau=1.0, single_precision=False)
system.lb = lbf

# None of these should work...
lbf[1,1,1].whatever_you_want = [1.0] * 19
lbf[1,:,1].whatever_you_want = [1.0] * 19

# this does not work, as expected
print(lbf[1,:,1].whatever_you_want)

[jngrad]

This is actually true for all script interface classes. In addition, the attribute is volatile, in the sense that it only exists on the script interface instance where it was set:

import espressomd
system = espressomd.System(box_l=[3, 3, 3])
wca=system.non_bonded_inter[0,0].wca
wca.my_value = 5
print(wca.my_value) # shows 5
print(system.non_bonded_inter[0,0].wca.my_value) # AttributeError

There are rare exceptions in custom classes, for example lbf.my_value and system.lb.my_value both refer to the same memory location.

Solving this issue requires a clever use of hasattr in the setter, such as:

--- a/src/python/espressomd/script_interface.pyx
+++ b/src/python/espressomd/script_interface.pyx
@@ -463,4 +463,9 @@ class ScriptInterfaceHelper(PScriptInterface):
 
     def __setattr__(self, attr, value):
+        if not hasattr(self, attr):
+            raise AttributeError(
+                f"Object '{self.__class__.__name__}' has no attribute '{attr}'; "
+                "ESPResSo script interface objects are thin wrappers for C++ "
+                "objects and cannot store extra data in user-defined attributes")
         if attr in self._valid_parameters():
             self.set_params(**{attr: value})

Then we would have to adjust a few edge cases in the script interface where we intentionally use volatile attributes. But the biggest change here would be requiring Cython 3.1. Previous versions of Cython implement the hasattr behavior from Python 2, which is known to be broken: instead of capturing AttributeError to detect the attribute absence, it captures all errors, even exit(1). MWE:

import espressomd.constraints
dubious_constraint = espressomd.constraints.ShapeBasedConstraint()
print(hasattr(dubious_constraint, "shape")) # raises RuntimeError in a .py file, prints False in a .pyx file

Here is the relevant line in the 3.1.0 alpha 1 changelog:

hasattr() now propagates exceptions that occur during lookup.

One could argue ShapeBasedConstraint should never have a null pointer in the first place, or at least return None, but then we would have to review all other classes for the same issue.

[jngrad]

There might be a way to pull this off with about 20 lines of extra Cython code to handle the null pointer thing and handle attributes defined via the @property decorator (involves an extra lookup in self.__class__.__dict__). Then a Cython upgrade is not required. With this change, 25 python test are failing, including all checkpoint tests, so more work is needed. I find this change useful, but I need to postpone it for now.