GPU BlockUniTensor Trace test uses wrong locator on rank-0 result (tests/gpu/BlockUniTensor_test.cpp)

[IvanaGyro]

Description

BlockUniTensorTest.gpu_Trace (renamed BlockUniTensorTest.GpuTrace on the reorganized codex/clean-up-test-files-and-reorganize-includes branch, same test body) throws in tests/gpu/BlockUniTensor_test.cpp. Reproduced on master@5d2d081 directly, so it predates and is unrelated to #1080 — this is a bug in the GPU test file itself, not in BlockUniTensor.

Reproduction

ctest --test-dir build/debug-openblas-cuda -R "BlockUniTensorTest.gpu_Trace" --output-on-failure

Observed error

[ RUN      ] BlockUniTensorTest.gpu_Trace
unknown file: Failure
C++ exception with description "
# Cytnx error occur at void cytnx::BlockUniTensor::_fx_locate_elem(cytnx::cytnx_int64&, std::vector<long unsigned int>&, const std::vector<long unsigned int>&) const
# error: [ERROR][BlockUniTensor][elem_exists] rank-0 scalar expects no locator indices.

# file : ../../src/BlockUniTensor.cpp (1384)" thrown in the test body.
[  FAILED  ] BlockUniTensorTest.gpu_Trace (308 ms)

Root cause

tests/gpu/BlockUniTensor_test.cpp traces a rank-2 UT_diag down to rank 0 and then indexes it with a non-empty locator:

tmp = UT_diag.Trace(0, 1);
...
EXPECT_DOUBLE_EQ(double(tmp.at({0}).real()), double(ans));   // wrong: tmp is rank-0
EXPECT_DOUBLE_EQ(double(tmp.at({0}).imag()), double(0));

The CPU counterpart in tests/BlockUniTensor_test.cpp (TEST_F(BlockUniTensorTest, Trace)) handles the same UT_diag.Trace(0, 1) result correctly, asserting rank-0-ness explicitly and using an empty locator, matching CLAUDE.md's "rank-0 scalar is not the same as shape {1}" guardrail:

tmp = UT_diag.Trace(0, 1);
...
EXPECT_EQ(tmp.rank(), 0);
EXPECT_TRUE(tmp.bonds().empty());
...
EXPECT_DOUBLE_EQ(double(tmp.at({}).real()), double(ans));
EXPECT_DOUBLE_EQ(double(tmp.at({}).imag()), double(0));

So the GPU test's tmp.at({0}) should be tmp.at({}) — the GPU test file just never picked up this correction that the CPU test file already has. BlockUniTensor::_fx_locate_elem is behaving correctly by rejecting the locator; this is a test bug, not a library bug.

Suggested fix

In tests/gpu/BlockUniTensor_test.cpp, change the two tmp.at({0}) calls following UT_diag.Trace(0, 1) to tmp.at({}), matching the CPU test's already-correct handling.

Environment

• CUDA 12.6.85 (conda cytnx env), GTX 1660 SUPER / RTX 2080 Ti.
• debug-openblas-cuda preset (ASan + RUN_TESTS=ON).