Fuzz-class fix: scale TranslationInvariance area tol with |translate|

With position-inclusive eps (InferEps uses bBox.Scale()), vertex snapping
at large translations changes area in either direction - not just shrinkage.
The prior EXPECT_LE was wrong; switching back to EXPECT_NEAR with a tol that
scales with |translate|: O(N * eps * edge_length) = O(|translate| * 1e-7),
so 1e-6 * (1 + rawArea + |tx| + |ty|) provides a 7.5x+ margin.

NumContour stays one-sided (EXPECT_LE): larger eps can only merge features,
never invent new contours.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: zmerlynn

test/cross_section_fuzz.cpp

@@ -941,11 +941,14 @@ void TranslationInvariance(const std::vector<double>& radii, double translateX,
   ExpectCrossSectionValid(shifted);
 
   // With position-inclusive eps (InferEps uses bBox.Scale()), far-from-origin
-  // polygons have larger eps, which can merge thin features. These are
-  // one-sided checks: eps merging can only shrink area and drop contours,
-  // never invent area or add contours.
-  const double tol = 1e-6 * (1.0 + std::fabs(rawArea));
-  EXPECT_LE(shifted.Area(), rawArea + tol);
+  // polygons have larger eps. Vertex snapping near the origin changes area by
+  // O(N * eps * edge_length) = O(|translate| * 1e-7), and the direction is
+  // not monotone (snapping can increase or decrease area). The tol scales with
+  // |translate| to cover this: at |t| = 1e3, tol ~ 2e-3; at 1e9, tol ~ 1e3.
+  // NumContour is one-sided: larger eps can only merge features, never invent.
+  const double tol = 1e-6 * (1.0 + std::fabs(rawArea) + std::fabs(translateX) +
+                             std::fabs(translateY));
+  EXPECT_NEAR(shifted.Area(), rawArea, tol);
   EXPECT_LE(shifted.NumContour(), base.NumContour());
 }