test: TestProject array_* builders accept a subscript over an undeclared dimension, so a fixture can pass on a model that never compiles

[bpowers]

Problem

TestProject's array_* builders (src/simlin-engine/src/test_common.rs) accept a subscript that names a dimension the fixture never declares, and nothing between the builder call and the compiler objects.

parse_array_declaration splits the bracket contents on commas and pushes the trimmed strings verbatim into Equation::ApplyToAll(dims, ..) / Equation::Arrayed(dims, ..); build_datamodel copies self.dimensions and self.variables into the Project without cross-checking them. So this builds without complaint:

TestProject::new("x")
    .indexed_dimension("D", 2)
    .array_aux("pp[D,2]", "1")   // no dimension named `2` exists

pp then fails to compile with BadDimensionName, and whatever the test asserts is asserted over a model that never compiled.

How it bit

On main, allocate_target_per_element_partials_decline in src/simlin-engine/src/db/ltm_rank_decline_tests.rs (line 275) built exactly that fixture. The test asserted that ALLOCATE's per-element LTM partials decline. They did "decline" -- because the whole model failed at BadDimensionName, not because the LTM decline logic fired. The test passed on a model that does not compile, and the behaviour it claimed to pin was never exercised.

PR #1057 (branch salsa-roofline-engine) fixed that one fixture (pp[d,xp] with indexed_dimension("xp", 4)) and added an assert_model_compiles guard to every test in that file. The builder itself is unchanged, so the same class recurs on the next fixture someone writes with a typo'd or forgotten dimension.

Why it matters

This is the failure mode the root CLAUDE.md rule "A test that hand-builds its inputs proves nothing about the inputs production supplies" is about: the fixture supplied a datamodel production can never produce (a variable dimensioned over a dimension that does not exist in the same project), and the test passed on an input that does not occur. The guard in ltm_rank_decline_tests.rs protects one file; there are ~2300 array_* call sites across ~60 test files in simlin-engine and libsimlin, none of which are protected.

Component

• src/simlin-engine/src/test_common.rs -- TestProject::{array_const, array_const_with_units, array_aux, array_stock, array_flow, array_with_ranges, array_flow_with_ranges, array_with_default_and_overrides, array_aux_direct, array_with_ranges_direct} and build_datamodel.

Proposed fix

A check in build_datamodel (not in each array_* method: builder chains declare dimensions and variables in arbitrary order, and build_datamodel is the single funnel every helper -- run_vm, error_diagnostics, flow_exprs, assert_compile_error_vm, ... -- goes through):

• For every variable whose equation is ApplyToAll(dims, ..) or Arrayed(dims, ..), every entry of dims must match a dimension in self.dimensions (covers indexed_dimension, indexed_subdimension, named_dimension*, and with_dimension).
• Match by canonicalize(name), not ==. The engine resolves dimension names canonically (variable::get_dimensions, GH engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541), and existing fixtures rely on it -- e.g. arrayed_helper_resolves_capitalized_dimension in tests/integration/ltm_array_agg.rs declares Region and the helper references it as region. An exact-match check would reject valid fixtures.
• Panic with the variable name and the offending dimension name, e.g. fixture "pp" is dimensioned over undeclared dimension "2" (declared: ["d"]).
• Skip when datamodel_override is set (from_datamodel wraps a converter-produced project; the builders are documented no-ops there).

About 20 lines. The cost is the sweep: this runs on every fixture in the workspace, so the change has to land with a cargo test --workspace pass that finds every fixture currently relying on the omission. A grep for tests that deliberately assert BadDimensionName through the builder found none -- the one test that wants an undeclared dimension (test_unknown_element_subscript_skips_unresolved_dimension in db/diagnostic_tests.rs) already builds its datamodel::Project directly via its own helper rather than through TestProject, which is the right shape for a fixture that needs an invalid model. If the sweep does turn up builder-based fixtures that need an undeclared dimension on purpose, they should move to that shape rather than get an opt-out flag on the builder.

Context

Identified during the PR #1057 review of ltm_rank_decline_tests.rs, where the hand-built fixture masked a compile failure as the LTM verdict the test was named for.