preconditioner="fmg" is inert on the semi-Lagrangian solvers: their solve() pre-runs setup and skips the base _build resolution

[lmoresi]

The base _build is where a solver's preconditioner choice is resolved against the mesh hierarchy (_apply_preconditioner_options), before snes.setFromOptions() reads the option bundle. The semi-Lagrangian solvers (AdvDiffusion, and the other solvers.py solvers on the same pattern) run the three setup stages directly inside their own solve():

if not self.is_setup:
    self._setup_pointwise_functions(verbose)
    self._setup_discretisation(verbose)
    self._setup_solver(verbose)

_setup_solver ends with self.is_setup = True, so the super().solve() that follows enters _build, hits if self.is_setup: return, and the resolution never runs. _pc_resolved stays False and nothing is recorded in pc_fallbacks.

Measured (feature/levelset-supg, amr-dev), a structured quad box with refinement=2 (three levels), AdvDiffusion with preconditioner = "fmg", one solve:

SLCN preconditioner='fmg' levels=3 pc=gamg pc_resolved=False fallbacks={}

The request is silently ignored: GAMG stays live, no warning, no record. preconditioner = "gamg" looks correct only because GAMG is what the scalar base pushes in __init__.

The Eulerian SUPG solver (#673) had the same pattern and the same symptom; it now builds through self._build(verbose) in place of the three calls, which is the one-line fix. The same change applies to each solve() in systems/solvers.py that pre-runs the stages (lines around 857, 1103, 1695 at 417f7b8). The pre-run exists so that update_pre_solve sees a built solver, but the history update needs the field, not the SNES, so the order can stay as it is.

Underworld development team with AI support from Claude Code