feat(language): a where may compare arithmetic over parameters on both lanes

[FBumann]

Prompt: Continue bumping to the head of #566

Note

The following content was generated by AI.

math-spec#566 lets either side of a where comparison be arithmetic over parameters — p_max > cost, 0.5 * p_max > 0, sum(p_max, over=generator) > 0. Both lanes build it now, so what the language accepts is what this package builds.

A side is compiled the way a constant side of a constraint is: the relational lane folds its constant fragments over the comparison's dims and joins the value column onto the carrier, the linopy lane evaluates it to a DataArray. A coordinate no fragment reaches stays null and the comparison reads false, which is what every other atom over a missing value does.

Two decisions to review rather than take on trust

Hard rule 0 moves. A mask now reads an expression, and a cases expression already read a mask, so the recursion is the language's own grammar and no ordering of a lane's two modules removes it. DELIBERATE_LAZY_IMPORTS was empty and that emptiness was the claim; it now holds two entries, one per lane, and docs/about/architecture.md says so. The alternative was threading an evaluator callback through masked() at seven call sites in two modules, which buys a parameter and a detour for every future reader.

NEVER_LOWERED excuses ArithmeticComparisonNode from both dispatch guards. It is in program.WhereNode, but lower_program rewrites every one into an ExpressionComparisonNode, so no lowered mask holds one and neither lane can dispatch on it. The exclusion is a claim about upstream, so test_no_never_lowered_node_survives_lowering checks it rather than trusting it, and the two assert_never pragmas name the same reason. Whether the union should carry a node lowering always replaces is a question for math-spec#566, and I have not raised it there.

Mutation table

tools.mutate, one guard, run before and after the probe that reaches it.

mutation	before the probe	after
the constants-only guard on a where side (predicates.py:223-225)	4023 passed — survived	caught

The guard survived because lower_program refuses a variable and a dual() on a where side before the engine is reached, so no model can carry one there. test_a_where_side_holding_a_variable_is_caught_rather_than_compiled_as_a_constant hands compile_predicate the node directly, which is the one caller that can reach it — without it the side would fold a variable's coefficient into a value column and mask on it.

Coverage moved

'p_max > cost' leaves the refusal sweep, where it asserted "compares two parameters", and joins ACCEPTED with '0.5 * p_max > 0' and 'sum(p_max, over=generator) > 0'. All three are swept by test_both_lanes_build_the_same_model for lane agreement on variable rows and termination status.

Gates

Run through uv rather than pixi, which is not installed in this environment — a departure from the pixi run check default. Re-run after the rebase.

gate	result
pytest -q -n auto	4024 passed, 251 skipped, 1 xfailed, 70.41s
ruff check .	All checks passed
ruff format --check .	326 files already formatted
pyrefly check	0 errors, 6 suppressed — was 4

Not run: test-bench, the sweep, and the docs build. The typst gallery cases skip for want of the binary, as they do in CI.

Deliberately not done: nothing was measured, so no number is claimed — a where side now compiles an expression where it used to read a column, and whether that costs anything on a real model is unmeasured. The stack is the base branch: this sits on #1696, which carries the pin.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NHXujoUG77G2SoDbhau5mS

---

Generated by Claude Code

[codspeed]

Merging this PR will not alter performance

✅ 24 untouched benchmarks
⏩ 58 skipped benchmarks¹

---

_{Comparing claude/happy-maxwell-9ldgqa-where-expressions (f23d814) with claude/happy-maxwell-9ldgqa-direction (4fc4b9e)}

Footnotes

1. 58 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

📚 lpspec | 🛠️ Build #34653174 | 📁 Comparing f23d814 against latest (556b558)

  🔍 Preview build  

14 files changed · ± 14 modified

± Modified

• about/architecture/index.html
• examples/piecewise_lp/index.html
• examples/pypsa_mixed_cycling/index.html
• examples/pypsa_spill/index.html
• examples/pypsa_storage/index.html
• examples/pypsa_store/index.html
• examples/pypsa_unit_commitment/index.html
• examples/pypsa_ladder/rung_02_storage/index.html
• examples/pypsa_ladder/rung_03_expansion/index.html
• examples/pypsa_ladder/rung_04_ramps/index.html
• and 4 more...