
ID	0029
Class	BUG
Severity	1
Status	ready

Models: —

Found: 2026-09-27, while diagnosing FHSIM-0028 against e0bbf4ea

Decision needed: —

Evidence

src/engine/integrator/sundials/FhIntegratorSundials.cpp:324-337: when the configured SUNDIALS linear solver is "SPGMR" and the model has a Jacobian system, linResolved is replaced by the name of the model's ResolveLinearSolverType() (DenseLU, SparseLU, BiCGSTAB, ...). :352 prints it as Linear solver : <linResolved>. For a DenseLU model that type is not used: SPGMR runs, SetupSpgmrPreconditioner returns no preconditioner for it (SundialsSolverBuilder.cpp:136-137), and :353 prints Preconditioner : none.

Observed on the fhsim_marine_elements S6 model (<Integrator Method="BDF">, no <LinearSolver>):

[Info]      Method          : cvodes/BDF
[Info]      Jacobian type   : dense
[Info]      Linear solver   : DenseLU
[Info]      Preconditioner  : none

SUNDIALS was running unpreconditioned SPGMR with maxl 5 (FHSIM-0028).

Effect

The log reads as a direct dense LU solve. A user looking for the reason BDF steps collapse is pointed away from the cause, because "DenseLU" with "Preconditioner: none" looks consistent.

Possible fix

Print the SUNDIALS linear solver that was created (SPGMR, DENSE, BAND), with maxl for SPGMR, and name the model's resolved type separately, e.g. Linear solver : SPGMR (maxl 5, unpreconditioned); model solver type DenseLU. Log a warning on the silent unpreconditioned path in SetupSpgmrPreconditioner (:136-137), as the other unpreconditioned paths already do. Log text only; no numerical change.

Test that would prove it

Run a dense-model BDF case without <LinearSolver> and assert that the log names SPGMR, not DenseLU, as the linear solver.

Risk

Low: log text only. A test or script that matches the old Linear solver : DenseLU line would need updating.