Fill: the height-field fill tries grids up to 64 spans

A cup's rim of seven or eight ellipse arcs meeting at corners was too wavy for
the height-field fill: its finest grid, 32 spans across, stayed 6.4e-5 off the
loop where the tolerance is 4.4e-5 (a smooth bicubic follows the height's kink
at a corner only as closely as its cells are small), and the loop went to the
general fill, seven faces where one fits. The grids now go to 64 spans (the
fit's system is banded); the first within tolerance is still the fill.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

crates/geom/src/fill.rs

@@ -16,7 +16,7 @@ use crate::bspline::{Dir, HPoint, NurbsSurface, Side, SplineError, SplineResult,
 /// Most loop points a fill takes.
 pub const MAX_LOOP_POINTS: usize = 8192;
 /// Most spans of the height field in each direction.
-pub const MAX_FILL_SPANS: usize = 32;
+pub const MAX_FILL_SPANS: usize = 64;
 /// How full a patch's middle may be made (`crown`): 0 the least-bending fill, positive fuller,
 /// negative flatter, in fractions of the loop's size. Outside this range the value is refused.
 pub const MIN_CROWN: f64 = -0.9;
@@ -89,7 +89,7 @@ pub struct LoopFill {
 ///   is free to turn at the rim (it is not the surface that made the rim). It does not depend
 ///   on the grid's directions. `λ = 10⁻⁷ ·` (fitted points) `/` (control points): just enough
 ///   to fix the inside, small enough not to pull the surface off the loop.
-/// - Grids of 4, 8, 16, 32 spans across the box's longer side are tried, with as many across
+/// - Grids of 4, 8, 16, 32, 64 spans across the box's longer side are tried, with as many across
 ///   the shorter side as keep the cells about square; a grid whose cells are under twice the
 ///   fitted points' largest step in a direction (cells between points would be left to the
 ///   energy alone) waits for denser sampling. The first grid within `tol` of every sampled