Ensure RandomizedSearchCV (and other optimizers) skips duplicated hyperparameter combinations

Describe the workflow you want to enable

RandomizedSearchCV and similar hyperparameter tuners need to handle duplicate hyperparameter combinations. This issue is particularly noticeable when a user has a small number of hyperparameters, especially when they are integers or categorical values.

Describe your proposed solution

A set of frozendict objects representing previously tried combinations should be maintained. When a new combination is generated (or retrieved), it should be skipped if it exists in this set, as running cross-validation on the same hyperparameters would be redundant.

Describe alternatives you've considered, if relevant

There are no such alternatives.

Additional context

I was surprised to learn that RandomizedSearchCV in sklearn 1.5.1 allows duplicate hyperparameters combinations.

model.cv_results_['params']

[{'binningprocess__max_n_prebins': 7},
{'binningprocess__max_n_prebins': 14},
{'binningprocess__max_n_prebins': 22},
{'binningprocess__max_n_prebins': 14},
{'binningprocess__max_n_prebins': 22},
{'binningprocess__max_n_prebins': 11},
{'binningprocess__max_n_prebins': 13},
{'binningprocess__max_n_prebins': 22},
{'binningprocess__max_n_prebins': 20},
{'binningprocess__max_n_prebins': 20}]

Start at RandomizedSearchCV and the related hyperparameter optimizer entry points; trace where sampled combinations are generated and represented in cv_results_['params']. Determine how duplicate combinations should be skipped, including integer and categorical spaces, and verify that the reported parameter combinations contain no repeats.