[DiadkoShmek]

When an arm has no observations for a metric that is measured on another arm, MergeRepeatedMeasurements still constructs an empty inverse-variance aggregation. This replaces the missing mean with zero and the SEM with infinity. Converting the transformed experiment data to observations then includes a metric that was never measured for that arm.

Skip empty arm/metric groups so their original missing values remain missing. Regression coverage checks either metric being absent, the weighted mean and variance of the measured metric, preservation of other arms, and a metric first observed in a later repeat.

Validation:

• Native Ax experiment/data extraction and observation conversion reproduce the failure on the base revision and pass with the change.
• python -m pytest -q ax/adapter/transforms/tests/test_merge_repeated_measurements_transform.py ax/adapter/transforms/tests/test_search_space_to_choice_transform.py ax/adapter/tests/test_discrete_adapter.py — 14 passed, 2 subtests passed.
• An independent exact-rational inverse-variance oracle passes 24 synthetic experiments, including reversed trial order (288 metric checks).
• Project-pinned ufmt and flake8 checks pass for both changed files.
• A separate public Experiment/Data check carried the result through TorchAdapter into actual BoTorch SupervisedDataset objects: the base fabricates three absent rows; the candidate retains the six measured rows, including a genuine zero.

AI assistance was used to identify, implement, and test this change.