Fix coordinate misalignment in expression merge

[FBumann]

Summary

Fixes silent data corruption when adding expressions whose coordinates have the same values but in different order (e.g. ['costs', 'Penalty'] vs ['Penalty', 'costs']). Previously, merge() used positional concatenation via join='override', ignoring labels entirely.

What changed

When merge() detects that two datasets share the same coordinate values in a different order, it now reindexes to the first dataset's order before concatenation. This only triggers when the coordinate value sets are identical — different subsets or different sizes are unaffected.

What is NOT affected

• Expressions with identical coordinates (same values, same order) — fast override path, unchanged
• Expressions with different dimension sizes — outer join path, unchanged
• Positional patterns like v.loc[:9] + v.loc[10:] — different value sets, override preserved

Bug reproduction script

import pandas as pd
import linopy

m = linopy.Model()
var = m.add_variables(coords=[pd.Index(['costs', 'Penalty'], name='effect')], name='var')
factors = pd.Series([2.0, 1.0], index=pd.Index(['Penalty', 'costs'], name='effect'))

result = var - 10 * factors

costs = float(result.const.sel(effect='costs'))
penalty = float(result.const.sel(effect='Penalty'))
print(f"costs: {costs} (expected -10), penalty: {penalty} (expected -20)")
assert costs == -10.0 and penalty == -20.0, "BUG: values misaligned!"

Benchmark script (dev-scripts/benchmark_merge_alignment.py)

"""Benchmark merge performance: compare override (same coords) vs outer join paths.

Run from repo root:
    python dev-scripts/benchmark_merge_alignment.py
"""

import statistics
import timeit

import pandas as pd

import linopy

ROUNDS = 10
CALLS_PER_ROUND = 10

m = linopy.Model()

time = pd.RangeIndex(8760, name="time")  # hourly year
space = pd.Index([f"node_{i}" for i in range(100)], name="space")
tech = pd.Index([f"tech_{i}" for i in range(20)], name="tech")

gen = m.add_variables(coords=[time, space, tech], name="gen")
cap = m.add_variables(coords=[space, tech], name="cap")

# Warm up
_ = 1 * gen


def bench(label: str, func, rounds: int = ROUNDS, number: int = CALLS_PER_ROUND):
    times = [
        timeit.timeit(func, number=number) / number * 1000 for _ in range(rounds)
    ]
    mean = statistics.mean(times)
    stdev = statistics.stdev(times)
    print(f"{label}: {mean:.1f} ± {stdev:.1f} ms  (n={rounds}x{number})")


# 1) Same-shape expressions (override path) — common in energy models
bench(
    "4-term sum (same shape, 8760x100x20)",
    lambda: 3.5 * gen + 1.2 * gen + 0.8 * gen + 2.0 * gen,
)

# 2) Broadcasting dimensions (gen has time, cap doesn't)
bench(
    "2-term sum (broadcasting, gen+cap)",
    lambda: 1 * gen + 1 * cap,
)

# 3) Many small additions
vars_list = [m.add_variables(coords=[time], name=f"v{i}") for i in range(50)]
bench(
    "50-term sum (1D, 8760)",
    lambda: sum(1 * v for v in vars_list),
    number=5,
)

Checklist

• Code changes are sufficiently documented
• Unit tests for new features were added
• A note for the release notes doc/release_notes.rst is included
• I consent to the release of this PR's code under the MIT license

[FBumann]

Im not entirely sure if this affects performance, but its at least worth raising i thought