refactor(api): specsolve no longer builds a linopy model, and its linopy extra and LaneError are gone

[FBumann]

Prompt: We should work on removing linopy entirely from lpspec. But keep it as a testing oracle!

Later, after CI failed: "Drop the xarray extra for now! Delete LaneError".

Note

The following content was generated by AI.

The linopy lane moves to tests/linopy_lane/ and stays the differential oracle. Nothing under src/ imports linopy. The package has no extra for linopy, pandas or xarray, and LaneError is deleted. No direct reference is left in the package metadata.

What changed, coverage moved, guards, gates, not done

The package

• specsolve.linopy, check(spec, sink='linopy'), lanes.LANES and lane_cannot_build_message are gone.
• The linopy extra is gone, and no xarray extra replaces it. to_pandas and to_dataarray import pandas and xarray lazily as before. When one is missing, the message says pip install pandas or pip install xarray. to_dataarray reads through pandas first, so an install with neither package is told pandas.
• allow-direct-references is gone: mathspec comes from PyPI (build(deps): specsolve depends on the mathspec 0.1.0 release from PyPI rather than a git tag #1754), and linopy is in a dependency group, which is not package metadata.
• LaneError is deleted. The engine's one construct that the language accepts and the engine cannot build (a constant beside a term, under an operator along a dimension, a constant summed beside a term is accepted by check and the linopy lane, refused by the engine #1137) now raises SpecsolveError. Its message names only the rewrite. The public surface drops from 28 names to 27, and architecture.md now says so.

The oracle

• tests/linopy_lane/ holds the eight modules, with their own CAPABILITIES table and a test-side OracleCannotBuildError(SpecsolveError).
• The dev group carries linopy @ git+…@master, xarray and pandas.
• pyrefly checks the oracle under strict, from project-includes, with search-path = ["src", "."].
• The PyPSA parity job takes linopy's pin from pyproject.toml with grep, so the pin has one home. Without it, pypsa pulled released linopy 0.9.1 into uv run's --with layer, and the oracle's import failed.

Coverage that moved

was	now
test_the_linopy_lane_stays_two_verbs	deleted: the oracle has no public surface
the _in_linopy_lane exemption in the two import fences	gone: the fences cover every module under src/
check(SPEC, sink='linopy') in test_quadratic_constraint	gone with the sink. The oracle's own refusal is still asserted
pytest.raises(LaneError) for the engine's gap (test_relational, test_compiler)	SpecsolveError, and still asserted not to be a LanguageError
LaneError from the oracle (test_corpus_parity xfail, test_conditioned_relations, test_linopy_lane, test_quadratic_constraint)	OracleCannotBuildError
the expression sweep's except LaneError	except OracleCannotBuildError, plus a SpecsolveError whose message says specsolve cannot build. Any other error is re-raised
bridge tests expecting pip install "specsolve[linopy]"	expect the missing package by name, and pandas where neither is installed

Guards, probed by hand after the commit: each mutation was made, the fence tests were run, the tree was restored with git checkout --, and __pycache__ was dropped on both sides.

mutation	result
import linopy at module level in src/specsolve/api.py	caught (2 fence tests)
import linopy inside check()	caught (the lazy-import fence)

Gates. pixi is not installable here, so the gates ran in uv environments:

• ruff check . and ruff format --check . pass.
• pyrefly check: 0 errors.
• Full environment (dev + docs groups, gurobi and xpress extras): pytest -n auto, 4139 passed, 250 skipped, 1 xfailed, 0 failed.
• A bare install (the built package, pytest, pytest-xdist and pyyaml; no linopy, xarray or pandas) on f65ad79: 2811 passed, 216 skipped. This is the run that reproduced the floors failure on the bridge test before the fix.
• zensical build --strict: "No issues found".
• The PyPSA parity job's command, run locally against mathspec v0.1.0: every rung matches, and the committed certificate is unchanged. CI's parity is green too.
• Not run: test-bench, and test-floors with its exact pinned floors.

Branch history: this branch merges #1751 after #1751 took in main (#1754, #1756, alpha.359). Another session had pushed the same merge, but its tree kept the [xarray] extra and the failing parity line. It is recorded with -s ours, so its content is superseded; nothing was force-pushed.

Not done: the internal docstrings that say "both lanes" (sources.py, assumptions.py, expressions.py) still hold, because the oracle enters through the same door.

Stacked on #1751, which adds ## Upcoming version.

🤖 Generated with Claude Code

https://claude.ai/code/session_016Pv7LzSgzt7Yn2K3ioyJXw

[codspeed]

Merging this PR will improve performance by 5.66%

⚡ 1 improved benchmark
✅ 23 untouched benchmarks
⏩ 82 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_window[commitment-s-specsolve-highs]	1.6 MB	1.6 MB	+5.66%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing claude/linopy-oracle-in-tests (f65ad79) with claude/confident-goodall-2iuw8n (bca5a67)}

Footnotes

1. 82 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[read-the-docs-community]

Documentation build overview

📚 specsolve | 🛠️ Build #34757422 | 📁 Comparing f65ad79 against latest (c3d74e4)

  🔍 Preview build  

9 files changed · ± 9 modified

± Modified

• index.html
• about/architecture/index.html
• about/changelog/index.html
• about/linopy/index.html
• about/roadmap/index.html
• examples/pypsa_ladder/index.html
• howto/debug/index.html
• reference/api/index.html
• reference/glossary/index.html