Time-dependent Dirichlet BCs are broken in solve(): essential values inserted at PETSC_MIN_REAL, mesh.t kernels see -1.8e308

[lmoresi]

Found by the adversarial review of the #407 fix (2026-07-24) — pre-existing, upstream of that change.

A Dirichlet BC that references mesh.t NaNs the solve:

poisson.add_dirichlet_bc(mesh.t, "Bottom")
poisson.solve(time=2.0)   # DIVERGED_FNORM_NAN; wall values -inf

Mechanism (read in PETSc source): the solve-side boundary insertion goes through DMPlexSNESComputeBoundaryFEM, which hard-codes time=PETSC_MIN_REAL into DMPlexInsertBoundaryValues; that value lands as petsc_t (-1.8e308) in the JIT'd essential-BC kernel, so any BC expression containing mesh.t evaluates to garbage. solve()'s docstring claim that 'expressions using mesh.t evaluate at this time' is empirically false for Dirichlet BCs.

Note the working counter-evidence: _assemble_volume_reaction(time=2.0) (post-#407) inserts g(t) correctly via a direct DMPlexInsertBoundaryValues call with the real time — so the kernels and the insertion machinery are fine; only the solve path's hard-coded PETSC_MIN_REAL is the problem. Fix direction: insert essential values explicitly with the solver's non-dimensional time before SNESSolve (or set the DM time in a way the PETSc path picks up — UW_DMSetTime exists but the BoundaryFEM path ignores it).

Related latent detail: boundary_flux() never forwards a solve time to _assemble_volume_reaction, so once this is fixed, CBF on a time-dependent wall would read g(0) — forward the time in the same change.

Underworld development team with AI support from Claude Code

[lmoresi]

Design ruling (Louis, 2026-07-24): retire the PETSc-time plumbing; time becomes a model-owned rampable constant

This failure is not news — the PETSc time-value route (UW_DMSetTime → petsc_t in kernels) was tried before and the trick failed; the PETSC_MIN_REAL insertion documented above is just where it visibly dies for essential BCs. The agreed direction instead:

1. t is a quantity, not a PETSc value. Implement time as a uw.quantity-style atom carried through the JIT constants[] block, so every compiled kernel (Dirichlet, Nitsche, natural, bodyforce, flux) reads the CURRENT time with no recompile per step — the same live-rampable mechanism as the δ-continuation constants.
2. Time lives on the MODEL, not the mesh. Simulations routinely carry multiple meshes (free-surface managers, dual meshes, adapt children), but they share one space-time — a per-mesh clock invites drift between them. The orchestration model already owns the reference scales, so it owns the clock; mesh.t stays as a resolving accessor for back-compatibility (all meshes report the model's time).
3. Once BC kernels read model time, the latent boundary_flux time-forwarding gap noted above disappears by construction — the reaction residual's kernels see the same clock as the solve did.

Prerequisite: #302. A constants[] atom that gets silently BAKED into the generated C source (the #302 bug) would freeze time inside kernels — the one failure mode this design cannot tolerate. #302 must be fixed (or at minimum, the time atom verified against its reproducer) before or together with this change.

Underworld development team with AI support from Claude Code