LP file export silently drops per-element bounds on binary variables, diverging from direct API

[FabianHofmann]

There is a discepancy betwenn io_api="direct" and io_api="lp" when in comes to bounds of binaries. The API should be consistent

Note

This issue was drafted with AI assistance (Claude Code); content reviewed before posting.

Version Checks (indicate both or one)

• I have confirmed this bug exists on the latest release of Linopy (v0.8.0).

Issue Description

Per-element bounds set on a binary variable are honored by the direct/matrix solver path but silently dropped by the LP file export, so the same model has different feasible sets depending on io_api.

• matrices._build_vars collects var.lower/var.upper for all variables, including binaries, and these lb/ub flow into the direct solver interfaces (e.g. gurobipy setAttr("LB"/"UB", ...)).
• io.bounds_to_file writes bound rows only for continuous + integers + semi_continuous, plus binaries with Variable.fixed == True (whole-variable .fix(), added in Fix variables via bound collapse instead of equality constraint #773). Binaries otherwise only appear in the binary section, which implies [0, 1] — any tighter per-element bounds (e.g. masking out a subset of entries with upper = 0) are silently lost.

This is longstanding behavior (not a 0.8.0 regression — before #773 binaries got no bound rows at all), but the divergence between the two paths is a silent correctness trap: a model that is correct with io_api="direct" relaxes when solved through an LP file.

Suggested fix: in bounds_to_file, also include binary variables whose bounds differ anywhere from (0, 1) — the LP format's bounds section may further restrict variables declared binary, and Gurobi/HiGHS honor this. Alternatively (or additionally), warn on LP export when a binary variable carries non-default bounds.

Reproducible Example

import pandas as pd
import linopy

m = linopy.Model()
x = m.add_variables(binary=True, coords=[pd.RangeIndex(4, name="t")], name="x")
x.upper = pd.Series([1, 1, 0, 0], index=pd.RangeIndex(4, name="t"))  # forbid x at t=2,3
m.add_constraints(x.sum() >= 2, name="atleast2")
m.add_objective(-1 * x.sum())

print(m.matrices.ub)  # [1 1 0 0] -> direct API honors the bounds, optimum -2

m.to_file("model.lp")
print(open("model.lp").read())
# LP file contains no bounds section at all; all four variables only appear
# under `binary` (implied [0,1]) -> LP-based solve returns optimum -4

Expected Behavior

Both solver paths see the same feasible set: the LP file should contain

bounds
0 <= x2 <= 0
0 <= x3 <= 0

(or the export should at least warn that non-default bounds on a binary variable are being dropped).

Installed Versions

Details

linopy 0.8.0, python 3.12, linux

[FBumann]

We should also allow to set binary bounds in add_variables(). It currently raises