fix: reject padded 2D layouts until execution handles strides

CompiledRaw2DNode::execute treats (batch, n0, n1) as tightly packed
and ignores idist/odist. The validation accepted idist >= n0*n1 which
would silently corrupt data for padded batches.

- Change idist/odist check from >= to == for rank-2 plans
- Add PaddedDistRejected test case

Ref: review comment on is_supported_2d_desc

Author: Artlesbol

ctest/test_plan.cpp

@@ -85,6 +85,19 @@ TEST(Plan2D, CustomStrideNotSupportedYet) {
   EXPECT_EQ(plan, nullptr);
 }
 
+TEST(Plan2D, PaddedDistRejected) {
+  // idist/odist > logical size must be rejected until 2D execution handles strides
+  int n[2] = {64, 32};
+  const int logical = n[0] * n[1];
+  flagfftHandle plan = nullptr;
+  EXPECT_EQ(flagfftPlanMany(&plan, 2, n, nullptr, 1, logical + 16, nullptr, 1, logical, FLAGFFT_C2C, 1),
+            FLAGFFT_NOT_SUPPORTED);
+  EXPECT_EQ(plan, nullptr);
+  EXPECT_EQ(flagfftPlanMany(&plan, 2, n, nullptr, 1, logical, nullptr, 1, logical + 16, FLAGFFT_C2C, 1),
+            FLAGFFT_NOT_SUPPORTED);
+  EXPECT_EQ(plan, nullptr);
+}
+
 TEST(Plan2D, InvalidParameters) {
   flagfftHandle plan = nullptr;
   EXPECT_EQ(flagfftPlan2d(&plan, 0, 32, FLAGFFT_C2C), FLAGFFT_INVALID_SIZE);

src/exec/c_api_internal.cpp

@@ -114,7 +114,10 @@ bool is_supported_2d_desc(const FlagFFTPlanDesc &desc) {
   const int64_t half_n1 = n1 / 2 + 1;
   const int64_t input_logical = real_inverse ? n0 * half_n1 : n0 * n1;
   const int64_t output_logical = real_forward ? n0 * half_n1 : n0 * n1;
-  if (desc.idist < input_logical || desc.odist < output_logical) {
+  // CompiledRaw2DNode::execute treats the entire (batch, n0, n1) block as
+  // tightly packed and ignores idist/odist.  Reject padded layouts until
+  // the 2D execution path plumbs strides through.
+  if (desc.idist != input_logical || desc.odist != output_logical) {
     return false;
   }
   return true;