fix: structural cheap_iszero hook in CLIL dropzeros! (OOM on symbolic values after #97)

[baggepinnen]

Follow-up to #97, implementing the fix proposed in this comment.

Problem

dropzeros!(::SparseMatrixCLIL) (added in #97) calls Base.iszero on the stored values. In get_new_mm the value type is Num, and iszero(::Num) is a semantic zero test (fraction_iszero → expand) that can be arbitrarily expensive. On MultibodyComponents.examples.suspension.HalfCar the process is OOM-killed inside dropzeros! at ~row 40 of a 295-row Num-valued mm (instrumented run; the pre-#97 commit completes fine).

Fix

Explicit stored zeros (as produced by sparsevec duplicate-index summation or alias cancellation) are always structural zeros, so a structural check suffices:

• StateSelection.CLIL.cheap_iszero(x) = iszero(x) — hook used by dropzeros!(::SparseMatrixCLIL) instead of Base.iszero,
• ModelingToolkitTearing overloads it for Num/SymbolicT via SU._iszero (structural Const(0) check, no expansion).

Numeric value types are unaffected (the default is Base.iszero).

Validation

• New regression test: a value type whose Base.iszero throws (standing in for "arbitrarily expensive") must still round-trip through dropzeros! correctly via the hook.
• The equivalent patch was validated on the HalfCar model: compiles without OOM, with behavior otherwise identical to pre-fix edge case in inline linear solve #97 (same emitted inline-linsolve blocks; details in the #97 comment).

🤖 Generated with Claude Code

[KristofferC]

Please cherry pick b281896 onto this branch

[baggepinnen]

Done — branch fast-forwarded to b281896 (its parent was the branch head, so this preserves your original commit/SHA rather than a re-applied copy). fbc/cheap-iszero-dropzeros tip is now b281896.

[AayushSabharwal]

Should be rendered unnecessary by JuliaSymbolics/Symbolics.jl#1893