fc3ae90-vv-campaign-promotions

• A local verification campaign measured seventeen subsystems against known answers and promoted them to supported: boundary conditions, the Picard and Explicit RK4 momentum solvers, the Poisson solver, periodic boundaries, domain definition, the Eulerian source, initial conditions, particle transport, the grid generator, monitoring, the post-processing pipeline, field statistics, spectra, the verification sources, case templates, and version management. Twenty-eight measurement records carry the evidence. The campaign found and fixed the defects below; several change results.
  • Solution-monitoring means and norms counted both ghost layers, so mean_speed read 1.7% and mean_ke 14% high on a periodic box. They now cover physical cells only.
  • Multi-rank Volume seeding gave each rank an equal count for an unequal subdomain, making particle density depend on the decomposition. Ranks now split the count by owned cells, and particle IDs are numbered contiguously across ranks.
  • The Poiseuille initial condition vanished at the first cell centres instead of on the walls and was applied across periodic axes; the parabolic inlet likewise vanished at the wall-adjacent cells and delivered 27% and 13% too little flux on 8 and 16 cells. Both now vanish on the walls and are uniform along periodic axes.
  • With dimensionalize: true, grid coordinates were scaled once per processed step, particle positions and velocities were not scaled, and Qcrit carried U_ref^2 rather than (U_ref/L_ref)^2. Each quantity is now scaled exactly once.
  • poisson_solver.method accepts only fgmres and cg: gmres, lgmres and bcgs converged their Krylov residual while the true residual stalled near 1e-3, and are now refused. Per-level max_it, rtol and atol reach the level solvers; they were emitted without PETSc's ksp_ prefix and ignored. poisson_solver.tolerance, which nothing read, is now refused.
  • A generated prescribed_flow inlet on a programmatic_c grid delivered roughly 2/n too much flux for n cells across (23% at 8, 12% at 16), because no grid was available to sample: it used uniform logical points and the continuous-area mean. It is now sampled on the bridge grid built from programmatic_settings and delivers the requested flux to 1e-9, stretched grids included, and a change to the grid settings rebuilds the profile asset. The boundary conditions subsystem is promoted with it.
  • Explicit RK4 stops at the step that exceeds its stability limit, naming the limit, instead of failing later in the Poisson solve with a message about multigrid depth. make smoke now runs it on a stable step and on an unstable one.
  • The solver refuses a left-handed grid. It used to negate metric vectors instead, which gave a mirrored duct forty times its bulk velocity. The grid generator refuses a transform list that ends left-handed, and a new reverse:axis=i|j|k transform renumbers one logical axis to fix it. Unknown transform, wall and path keys are refused instead of silently ignored.
  • --statistics-state carry stages the restart bundle even for an analytical Eulerian source; it used to fail at start-up. summarize --plot-spectrum finds spectra that post recipes write into their own subdirectory.
  • The PETSc -info log survives the fresh-run log wipe; the reopen no longer leaks its filename. A profiling function name that matches no instrumented function draws a warning. versions activate -- -j8 passes -j8 to make instead of handing it to Git.
  • Documentation corrections: the convergence modes only log and never judge or stop a run; CornerAveraged interpolation is below second order even on uniform grids (order 1.68 in L2); TGV3D is a prescribed kinematic field, not a Navier-Stokes solution; the periodic wall-bounded pseudo-time stall no longer reproduces; the IEM scalar is inert because nothing seeds a non-zero Psi; and an external input reference's registration checksum is recorded but not compared, so a changed target is used as it now is, under a new asset identity.
  • New: scalar_transport.iem_constant sets the IEM mixing constant C_IEM, default 2.0, for sensitivity studies. Particle scalar mixing stays experimental: nothing seeds a non-zero Psi yet.
  • A supported subsystem may now be demoted to experimental, with a recorded demotion_reason saying what the earlier claim did not establish.