[BUG] CI Failure on ANN Descent

While CI was running on a Java PR, this C++ test failed:

[ RUN      ] AnnNNDescentTest/AnnNNDescentTestI8_U32.AnnNNDescent/116
[859479][06:35:48:642681][info  ] Recall = 0.885602 (113357/128000), the error is 14.4% above the threshold (eps = 0.001000).
/tmp/conda-bld-output/bld/rattler-build_libcuvs/work/cpp/tests/neighbors/ann_nn_descent/../ann_nn_descent.cuh:151: Failure
Value of: eval_neighbours(indices_naive, indices_NNDescent, distances_naive, distances_NNDescent, ps.n_rows, ps.graph_degree, 0.001, min_recall)
  Actual: false (actual recall (0.88560156249999999) is lower than the minimum expected recall (0.90000000000000002); eps = 0.001. )
Expected: true

[  FAILED  ] AnnNNDescentTest/AnnNNDescentTestI8_U32.AnnNNDescent/116, where GetParam() = dataset shape=2000x1024, graph_degree=64, metric=6, device
 (87 ms)

https://github.com/rapidsai/cuvs/actions/runs/18864158066/job/53829943019?pr=1469

Start with the failing AnnNNDescent test and the assertion at cpp/tests/neighbors/ann_nn_descent/../ann_nn_descent.cuh:151, then inspect the linked CI job for the Java PR. Reproduce the dataset shape=2000x1024 case and investigate why recall is 0.8856 instead of the required 0.9000; done means the test passes its minimum recall check.