Symbolics NaNMath.pow lowering variants break builder sanitizers

[finmod]

mtk_issue_bundle.zip

Title: Symbolics codegen emits varied NaNMath.pow lowered forms that escape sanitizers and break compiled builder functions

Summary

When generating numeric builder functions, Symbolics/ModelingToolkit sometimes lowers fractional powers (x^α) into varied textual/AST variants that reference NaNMath.pow in dotted/wrapped/parenthesized shapes. These shapes escape robust string/AST sanitizers in our initialization pipeline and survive into the final lowered/compiled residuals, causing runtime MethodError when the compiled builders run.

Environment

• OS: Windows
• Repo: DyadDemos MacroeconomicModelDemo (local)
• Julia project: --project=., attached Project.toml/Manifest.toml in repo root
• Files attached: failed_builder_exprs.log, init_solver_clean.jl, check_consistency.jl, run_cell14.jl, and a zipped bundle of these files

Reproduction (minimal steps)

1. Clone the attached repo (or use provided bundle).
2. From project root run:

julia --project=. check_consistency.jl

3. Observe runtime builder lowering printed in generated/failed_builder_exprs.log and MethodError during compiled residual smoke-test.

Representative failing lowered node forms (observed in logs)

• (NaNMath.pow)(ˍ₋arg1[40], 0.6)
• (NaNMath.pow)((+)(1.0e-8, (abs)(ˍ₋arg1[40])), 0.6)
• (NaNMath.pow)((+)(ˍ₋arg1[63], (*)(-1, ˍ₋arg1[48])), 0.5)
• (NaNMath.pow)((/)((+)(1, ˍ₋arg1[59]), (+)(1, ˍ₋arg1[63])), 2.0)

What I tried locally

• Implemented multi-stage sanitizers in src/init_solver_clean.jl (string-level guarded replacers, regex normalizers, AST walker, deterministic prewalk, placeholder-based textual→parse→AST fallback).
• Added _safe_pow and a Main.NaNMath shim so lowered NaNMath.pow resolves to a safe fallback.
• Despite these, lowered builder outputs still include NaNMath.pow in the shapes above and cause a runtime MethodError in the smoke test.

Suspected cause / pointers

• Codegen lowering for fractional power ^ appears to emit multiple non-canonical call forms (parenthesized/dotted/wrapped), which make downstream sanitizers fragile.
• Suggest making the lowering canonical (consistent Expr form) or exposing a hook so downstream code can reliably canonicalize pow lowering.

Attachments

• failed_builder_exprs.log (full run output showing lowered builder functions)
• init_solver_clean.jl (our sanitizer + shim implementations)
• check_consistency.jl (driver to reproduce)
• run_cell14.jl (short script used to reproduce notebook cell warnings)
• zipped archive: mtk_issue_bundle.zip

Notes

• I can prepare a smaller minimal reproduction if the maintainers prefer a reduced example. I am available to iterate on narrow testcases or propose a small patch to canonicalize lowering.

[AayushSabharwal]

I don't follow - what's the problem with NaNMath.pow? All the examples above are valid syntax?

[ChrisRackauckas]

Yeah this needs a simple MWE. Those generated codes look perfectly fine.

[finmod]

Update: We installed the missing Optimization and OptimizationOptimJL packages (v4.8.0 and v0.4.5). After this:

✅ All system variables have initial conditions (41/41)
✅ Solver runs without MethodError
✅ No NaNMath.pow runtime issues detected
✅ Dyad compilation works correctly

Finding: The issue either was resolved by the package updates or is being handled gracefully by the current environment. Our complex SFC model compiles and integrates with MTK without errors.

For a definitive fix in Symbolics, your request for a minimal reproducible example still stands.