`set_normalized_rhs` changes LessEqual constraint sense to equality in HiGHS model

[AllisterFrazier]

Hi, thanks for maintaining PyOptInterface.

I may be misunderstanding the intended semantics of set_normalized_rhs, but I found a case where calling it on a LessEqual linear constraint appears to change the constraint into an equality constraint.

Environment

• pyoptinterface: 0.6.1
• solver: HiGHS
• Python: 3.13
• OS: Windows

Minimal reproducer

from pathlib import Path

import pyoptinterface as poi
from pyoptinterface import highs

before = Path("rhs_before.lp")
after = Path("rhs_after.lp")

m = highs.Model()

x = m.add_variable(name="x", lb=0, ub=10)

m.set_obj_sense(poi.ObjectiveSense.Minimize)
m.set_objective_coefficient(x, 1.0)

expr = poi.ScalarAffineFunction()
expr.add_term(x, 1.0)

c = m.add_linear_constraint(
    expr,
    sense=poi.ConstraintSense.LessEqual,
    rhs=5.0,
    name="limit",
)

m.write(str(before))
m.optimize()
print("before RHS change:", m.get_variable_attribute(x, poi.VariableAttribute.Value))

m.set_normalized_rhs(c, 4.0)

m.write(str(after))
m.optimize()
print("after RHS change:", m.get_variable_attribute(x, poi.VariableAttribute.Value))

print("before LP row:")
print(next(line.strip() for line in before.read_text().splitlines() if line.strip().startswith("limit:")))

print("after LP row:")
print(next(line.strip() for line in after.read_text().splitlines() if line.strip().startswith("limit:")))

Observed Behavior

before RHS change: 0.0
after RHS change: 4.0

before LP row:
limit: +1 x <= +5

after LP row:
limit: +1 x = +4

The constraint in the written LP file changes from inequality to equality. The solve result also changes: since this is a minimization model with x >= 0, a preserved x <= 4 constraint should still allow x = 0. Instead, the model behaves like x = 4.

Expected behavior

I expected set_normalized_rhs(c, 4.0) to preserve the original constraint sense and update only the RHS, resulting in:

limit: +1 x <= +4

If this behavior is intentional because set_normalized_rhs operates on some canonical equality representation, is there a recommended API for changing the RHS of an existing inequality constraint while preserving its sense?

[AllisterFrazier]

Using a commercial license, I was able to verify that the expected behavior is observed when using Gurobi.
So perhaps this is actually a HiGHS bug?

Here is the snippet:

from pathlib import Path

import pyoptinterface as poi
from pyoptinterface import gurobi

before = Path("rhs_before.lp")
after = Path("rhs_after.lp")

m = gurobi.Model()

x = m.add_variable(name="x", lb=0, ub=10)

m.set_model_attribute(
    poi.ModelAttribute.ObjectiveSense,
    poi.ObjectiveSense.Minimize,
)
m.set_objective_coefficient(x, 1.0)

expr = poi.ScalarAffineFunction()
expr.add_term(x, 1.0)

c = m.add_linear_constraint(
    expr,
    sense=poi.ConstraintSense.LessEqual,
    rhs=5.0,
    name="limit",
)

m.write(str(before))
m.optimize()
print("before RHS change:", m.get_variable_attribute(x, poi.VariableAttribute.Value))

m.set_normalized_rhs(c, 4.0)

m.write(str(after))
m.optimize()
print("after RHS change:", m.get_variable_attribute(x, poi.VariableAttribute.Value))

print("before LP row:")
print(next(line.strip() for line in before.read_text().splitlines() if line.strip().startswith("limit:")))

print("after LP row:")
print(next(line.strip() for line in after.read_text().splitlines() if line.strip().startswith("limit:")))

And the result:

after RHS change: 0.0
before LP row:
limit: x <= 5
after LP row:
limit: x <= 4