read_netcdf sorts containers alphabetically, permuting matrices.A after a round trip

[FabianHofmann]

Note

The following content was generated by AI (Claude Code) while reviewing #929, verified by @FabianHofmann.

read_netcdf rebuilds the variable, expression and constraint containers with sorted(...) over the names (linopy/io.py, read_netcdf), so a model comes back in alphabetical instead of insertion order. Labels are preserved, but matrices.A orders its columns by container iteration, so the matrix of a round-tripped model is a column permutation of the original. Same for the constraint rows. Independent of the semantics option and of frozen constraints.

Repro

import linopy
import numpy as np

m = linopy.Model()
z = m.add_variables(lower=0, name="z")
a = m.add_variables(lower=0, name="a")
m.add_constraints(z + a >= 10, name="con1")
m.add_constraints(z - a <= 5, name="con2")
m.add_objective(z + 2 * a)

m.to_netcdf("model.nc")
m2 = linopy.io.read_netcdf("model.nc")

print(list(m.variables), list(m2.variables))
print(np.array_equal(m.matrices.A.toarray(), m2.matrices.A.toarray()))

['z', 'a'] ['a', 'z']
False

Proposal

Store the container order in the file (e.g. an attribute listing the names) and rebuild in that order, falling back to the sorted order for files without it.