Fix: reject invalid inverse lens-map weights

Check spherical area weight sums against a fixed 1e-8 tolerance and normalize accepted weights before interpolation. Skip invalid images in both catalog paths.

Add a thin-triangle regression for both parities, document the measured tolerance margin, and refresh HDR fixtures with strict comparisons restored.

src/frame.c

@@ -930,6 +930,17 @@ static int spherical_barycentric_weights(const double point[3], const double a[3
   weights[0] = spherical_area(point, b, c) / area;
   weights[1] = spherical_area(point, c, a) / area;
   weights[2] = spherical_area(point, a, b) / area;
+  /* A thin source triangle can admit an exterior point through the tolerant
+   * edge test. Unsigned subareas then do not partition the total area, and
+   * interpolating with their ratios can move an image outside its triangle.
+   * Reject that case before normalizing roundoff in valid convex weights.
+   * The 1e-8 bound is dimensionless; see the inverse-map design notes. */
+  const double weight_sum = weights[0] + weights[1] + weights[2];
+  if (!isfinite(weight_sum) || weight_sum <= 0.0 ||
+      fabs(weight_sum - 1.0) > 1e-8)
+    return -1;
+  for (int i = 0; i < 3; ++i)
+    weights[i] /= weight_sum;
   return 0;
 }
 
@@ -1055,7 +1066,7 @@ static int splat_catalog_tile(const Star *stars, size_t count,
                                         context->vertex[0]->n_infinity,
                                         context->vertex[1]->n_infinity,
                                         context->vertex[2]->n_infinity, weights))
-        return -1;
+        continue;
     }
     if (!owns_source_boundary(context->triangle, weights))
       continue;