test: pin alignment contract for variable/expression/constraint bugs

[MaykThewessen]

Summary

Adds test/test_alignment_contract.py with six xfail(strict=True) regression
tests that pin the intended coordinate-alignment contract reported across the
open alignment-bug cluster (#586, #706, #707, #708, #709).

The bugs all describe the same root issue: linopy's operators and
add_variables handle coordinate alignment inconsistently — some by label,
some positionally, some silently dropping dims. The discussion on each issue
keeps re-litigating what the contract should be. A failing test pins the
answer: every operator aligns by label, add_variables preserves coords
order and shape regardless of bound type.

Issue → test mapping

Issue	Test	Contract pinned
#707	test_subtract_then_le_matches_direct_le	x - a <= 0 and x <= a build the same constraint
#708	test_add_aligns_by_label	x + a aligns by label, not by position
#708	test_subtract_aligns_by_label	x - a aligns by label
#708	test_divide_aligns_by_label	x / a aligns by label and x / a == x * (1 / a)
#706	test_add_variables_dataarray_bounds_preserve_coords_order	add_variables dim order follows coords for DataArray bounds with missing dims
#709	test_add_variables_pandas_bounds_broadcast_missing_dim	add_variables broadcasts pandas bounds to full coords shape

#586 is intentionally not included as its own test. Its original repro
(m.add_constraints(x >= series) with reordered labels) no longer reproduces
on master — the add_constraints path now label-aligns correctly. #707's
description already notes this: "#586's own snippet … takes the label-aligned
path and no longer reproduces, but the bug is alive in arithmetic." The
surviving arithmetic piece is covered by the #707 / #708 tests above, so
#586 is effectively merged into that coverage.

Why xfail strict=True

Each test asserts the expected behaviour (the bug fixed), not the current
buggy output. As long as the bug reproduces the test xfails; the moment the
fix lands the test starts passing, strict=True turns the XPASS into a CI
failure, and the maintainer is forced to remove the xfail mark in the same
commit that closes the issue. That is the right state machine: the test
flips from "bug exists" to "regression guard" in one atomic step.

Confirmed by running locally against master (commit 37af4ba):

test/test_alignment_contract.py xxxxxx                                  [100%]
============================== 6 xfailed in 0.45s ===============================

Notes for maintainers

• Please push back on any of the "expected behaviour" assertions before merge
  — pinning the contract is the point of this PR, so disagreement here is
  exactly the conversation we need. In particular: the divide test pins both
  "/ aligns by label" and "x / a == x * (1 / a)". If you'd rather not
  commit to the algebraic identity, that line is easy to drop.
• The tests are deliberately small (≤30 lines each), share two minimal
  fixtures, and use only pd.Series / xr.DataArray — no solver call, no
  network I/O. They run in well under a second.
• If you'd prefer this content to live inside test/test_constraints.py's
  existing TestConstraintCoordinateAlignment class or alongside
  test/test_variables.py, happy to relocate — kept it in a dedicated file
  to make the contract-pinning intent obvious.

Related issues: #586, #706, #707, #708, #709 (and #591, which proposes the
strict-alignment convention these tests would lock in).

[FBumann]

@MaykThewessen thanks for the PR. Ill add a bunch of similar tests soon. Could you try to cover more datatypes etc with pytest.parameterize, while keeping the test themselves and the parameters readable?

[MaykThewessen]

Thanks @FBumann — pushed an update (force-with-lease, ce0db80 → 8ae16d7) that parametrizes the six tests over labelled coefficient / bound containers and dtypes. 6 cases → 21.

Coverage matrix:

Test	Issue	Parametrize over
test_subtract_then_le_matches_direct_le	#707	pd.Series, xr.DataArray × float64, int64 (4)
test_add_aligns_by_label	#708	same (4)
test_subtract_aligns_by_label	#708	same (4)
test_divide_aligns_by_label	#708	same (4)
test_add_variables_dataarray_bounds_preserve_coords_order	#706	xr.DataArray × float64, int64 (2)
test_add_variables_pandas_bounds_broadcast_missing_dim	#709	pd.Series × float64, int64, pd.DataFrame × float64 (3)

All 21 parameter cases still xfail(strict=True) against master — confirmed locally:

test/test_alignment_contract.py ......x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x...x [100%]
============================= 21 xfailed in 1.29s ==============================

Readability choices:

• Shared module-level constants (_COEF_LABELS, _VAR_LABELS, _X, _Y) so each parameter case is built by a one-line helper (_series("float64"), _dataarray("int64"), …) — the parameter IDs read as series-float64 / dataarray-int64 / dataframe-float64 rather than the auto-generated repr soup.
• xfail mark stays on the test, not on each parameter case — every dtype/container combination reproduces the bug today, so a per-case mark would just be six identical reasons. The moment the upstream fix lands, all cases for a given test will XPASS together and the strict-mark forces removal in the same commit.
• Unlabelled inputs (np.ndarray, list, scalar) are intentionally not parametrized — they have no labels for the contract to align against, so positional behaviour is correct for them and they would silently turn into XPASS noise. I left a note in the module docstring explaining this so it's not re-litigated in review.

Happy to widen further (e.g. xr.Dataset, ND DataArray, mixed dtypes via complex128 / Int64 extension) if you'd like — wasn't sure where the readability budget ran out. Also happy to fold these into your follow-up suite if you'd rather keep this PR minimal and land the broader parametrization in your own commit.