TTM folds grids on realistic bathymetry

Thompson-Thames-Mastin elliptic generation on a Gaussian seamount
(`D(x) = L*(-1 + a exp(-b(x-x_c)^2))`, `L*` = 1000 m, `a` = 0.5, `b` = 8),
99 x 38, with the boundary curves in `grids/seamount/`:

```
iters Jacobian min J
5 MIXED (9 neg) -1.882e+04
10 MIXED (72 neg) -1.882e+04
25 MIXED (420 neg) -1.772e+04
50 MIXED (866 neg) -1.141e+04
100 MIXED (1005 neg) -3517
300 MIXED (890 neg) -0.7698
TFI, same curves : ALL + (usable)
```

The xi lines collapse into a singular fan near the surface above the crest. This
is not slow degradation from over-iteration — it is folded within 5 iterations.
TFI on identical boundary curves is clean.

A folded grid is a *locally* left-handed grid. `jacobian2D` returns the signed
determinant, `grad2DCurv` divides by it, and nothing checks. A run on that grid
would produce finite, plausible output with the metric tensor inverted in a
patch directly over the interesting bathymetry.