e6a0696-les-grid-diagnosis-fixes

• Diagnosing the Humphrey bent-channel failure found and fixed defects in the LES filter widths, the Clark gradient term and the metric helpers, added the Vreman and WALE subgrid models and the Scotti filter width, and made the grid report state what predicts grid-scale oscillation.
  • The failure itself was numerical, not a solver or physics fault: central convection, which any LES model selects regardless of central_diff, leaves a two-cell streamwise oscillation undamped where the streamwise cell Reynolds number is large and the subgrid model supplies almost nothing, as in a laminar or transitional stretch. Solver reference page 08 now says so.
  • The Clark gradient term multiplied per-cell velocity differences, which already carry the cell spacing, by a squared cell extent as well. Its stress scaled as Delta^4 and was smaller than intended by Delta^2 - 1e-4 to 1e-6 on a production mesh - so any earlier run with gradient_model.enabled effectively ran without it. It now matches (Delta^2/12) du/dx du/dx, and such runs will behave differently.
  • The geometric_mean and max_edge filter widths used the Cartesian components of the cell diagonal, so an identical cell got up to nine times the eddy viscosity once a bend turned it. They now use the cell's extents along its own grid directions, and their widths change on curved grids.
  • A fixed 1e-12 floor on the determinant of the face-area vectors declared any wall-resolved cell singular, and the refusal was dropped, so selecting max_edge or geometric_mean on such a grid produced a NaN eddy viscosity. Degeneracy is now judged against the cell's own scale and the error propagates. The three helpers involved have no production caller left and are marked deprecated.
  • New: filter_width: scotti (Scotti, Meneveau and Lilly 1993), and the coefficient-free models model: vreman and model: wale, which vanish in pure shear and need no test filter, averaging or coefficient field. Vreman weights each grid direction by the cell's own edge. All are experimental and unvalidated against a reference flow.
  • The constant model accepts filter_width directly instead of through a PETSc passthrough. Each LES model now refuses the parameters of the others, including dynamic_frequency under the constant model, and -les is range-checked instead of being read into an uninitialized value when absent.
  • The grid generator report states the wall-unit length and what it was built from, since an Re_tau quoted on a half-width with a full-width length_ref halves every plus number, and adds the cell Reynolds number, the effective cell Reynolds number under an assumed --nut-ratio, and the time viscosity takes to clear a two-cell oscillation along each direction.
  • Every tracked C source and header now routes to a freshness surface or contract; 41 of 75 routed nowhere. The test map gave src/les.c no unit-les target.