`remove_variables` deletes unrelated constraints: `-1` means "masked" in a variable but "empty term slot" in a constraint

[YassineAbdelouadoud]

Version Checks (indicate both or one)

• I have confirmed this bug exists on the lastest release of Linopy.

• I have confirmed this bug exists on the current master branch of Linopy.

Issue Description

Summary

Since 0.8, Model.remove_variables(name) deletes every constraint for which Constraint.has_variable(variable) is true. That test is

# linopy/constraints.py
def has_variable(self, variable: variables.Variable) -> bool:
    return bool(self.data["vars"].isin(variable.labels.values.ravel()).any())

-1 appears on both sides of that comparison with two unrelated meanings:

• in variable.labels, -1 marks an entry masked out by mask=;
• in constraint.data["vars"], -1 marks an unused term slot in the padded _term dimension.

So any masked variable tests as "used by" any constraint that has a padded term slot, and removing the former silently deletes the latter. Models that use mask= at all are affected, and the more masking there is, the more constraints are destroyed.

The failure is quiet: the model still builds, it has simply lost constraints, so it typically goes on to solve unbounded or to a wrong optimum. The emitted UserWarning names constraints that genuinely have nothing to do with the removed variable, which makes it read like a false alarm rather than a symptom.

Reproducible Example

import pandas as pd
import xarray as xr
import linopy
from linopy import Model

print("linopy", linopy.__version__)

idx = pd.Index([0, 1, 2], name="i")
m = Model()

a = m.add_variables(
    name="a", lower=0, coords=[idx],
    mask=xr.DataArray([True, False, True], coords=[idx]),
)
b = m.add_variables(
    name="b", lower=0, coords=[idx],
    mask=xr.DataArray([True, True, False], coords=[idx]),
)

m.add_constraints(b >= 1, name="only_b")

print("a labels   :", a.labels.values)                                  # [ 0 -1  2]
print("b labels   :", b.labels.values)                                  # [ 3  4 -1]
print("only_b vars:", m.constraints["only_b"].data["vars"].values.ravel())  # [ 3  4 -1]

m.remove_variables("a")
print("only_b still in model:", "only_b" in m.constraints)

Expected Behavior

only_b should survive: it does not use a. More generally, has_variable should ignore the -1 sentinel on both sides.

Suggested fix

Restrict the membership test to real labels:

def has_variable(self, variable: variables.Variable) -> bool:
    labels = variable.labels.values.ravel()
    labels = labels[labels >= 0]
    used = self.data["vars"]
    return bool(used.where(used >= 0).isin(labels).any())

(or filter used to >= 0 first — the point is that neither side should let -1 match -1.)

Installed Versions

Details

- linopy 0.7.0 (correct), 0.8.0 and 0.9.0 (both affected) - Python 3.13, xarray 2025.12.0, pandas 2.x

[FabianHofmann]

Thanks for the issue which is high priority. Happy to review a PR otherwise I will tackle it next week.