This is arising from #1026, TSC refactoring to the new fit/_fit model, which causes orchestration tests to break post-refactor, in test_stat in test_orchestration.py.

There is a ProximityForest used in the current orchestration test, which fails after the refactor. If it is switched out for basically any other time series classifier, the orchestration test passes. A medium-depth look at the tracebeek seems to suggest that it is some unexpected behaviour of ProximityForest rather than of the orchestration framework.

To reproduce the bug: clone #1026, but revert test_orchestration.test_stat to the pre-PR state.
Run test_orchestration.test_stat.