Wrong metadata sharing during fit with OptunaSearchCV

[Kreol64]

Expected behavior

1. While running OptunaSearchCV and enable_metadata_routing=True, set_score_request() is expected to share metadata with scoring function only and set_fit_request() is expected to share metadata with fit function only
2. If error_score='raise', any failed Trial is expected to raise error

Environment

• Optuna version: 4.1.0
• Optuna Integration version: 4.2.0
• Python version: 3.10.8
• OS: Windows-10-10.0.19045-SP0
• (Optional) Other libraries and their versions:
  sklearn 1.4.1.post1

Error messages, stack traces, or logs

1)  UserWarning: Failed to report cross validation scores for TerminatorCallback, with error: The length of `scores` is expected to be greater than one.
  warnings.warn(warn_msg)

2) TypeError: Ridge.fit() got an unexpected keyword argument 'score_param'

Steps to reproduce

1. If refit=False and n_jobs=1 then just observe warning that (I believe) should not be there
2. If refit=True or n_jobs>1 observe incorrect passing of score-only parameters into fit function
3. If metric returns NaN, error should be raised if error_score='raise". Currently a trial is failed and error message is empty - very hard to debug.

import numpy as np
import optuna
import sklearn
from sklearn.linear_model import Ridge
from sklearn.metrics import make_scorer
from optuna.integration import OptunaSearchCV
from sklearn.model_selection import GridSearchCV

sklearn.set_config(enable_metadata_routing=True)


def custom_metric(y: np.ndarray, y_pred: np.ndarray, sample_weight: np.ndarray, score_param: str, *args, **kwargs) -> float:
    assert score_param == 'a'
    assert np.all(sample_weight == np.ones(4))
    return 0.01


x = np.random.randn(10, 2)
y = np.random.randn(10)
y[-1] = np.nan
w = np.ones(len(y))
score_param = 'a'

sklearn_grid_params = dict(
    param_grid={'alpha': [0.0, 0.1]},
    scoring=make_scorer(
        custom_metric,
        greater_is_better=True,
    ).set_score_request(
        sample_weight=True,
        score_param=True
    ),
    cv=[(np.arange(10)[:6], np.arange(10)[6:])],
    return_train_score=False,
    refit=False
)

optuna_grid_params = dict(
    scoring=make_scorer(
        custom_metric,
        greater_is_better=True,
    ).set_score_request(
        sample_weight=True,
        score_param=True
    ),
    n_jobs=1,
    error_score='raise',
    n_trials=10,
    cv=[(np.arange(10)[:6], np.arange(10)[6:])],
    param_distributions={'alpha': optuna.distributions.CategoricalDistribution([0.0, 0.1])},
    random_state=77,
    return_train_score=False,
    refit=False
)

sklearn_grid = GridSearchCV(
    estimator=Ridge().set_fit_request(sample_weight=True),
    **sklearn_grid_params
)

optuna_grid = OptunaSearchCV(
    estimator=Ridge().set_fit_request(sample_weight=True),
    **optuna_grid_params
)

# runs without issues as expected. score_param is only fed into scoring function
sklearn_grid.fit(X=x, y=y, sample_weight=w, score_param=score_param)

# Observe warning "UserWarning: Failed to report cross validation scores for TerminatorCallback, with error: The length of `scores` is expected to be greater than one."
optuna_grid.fit(X=x, y=y, sample_weight=w, score_param=score_param)

# Observe error when refit=True: TypeError: Ridge.fit() got an unexpected keyword argument 'score_param'
optuna_grid_params['refit'] = True
optuna_grid = OptunaSearchCV(
    estimator=Ridge().set_fit_request(sample_weight=True),
    **optuna_grid_params
)
optuna_grid.fit(X=x, y=y, sample_weight=w, score_param=score_param)

# Observe same error when refit is set back to false and number of n_jobs > 1
optuna_grid_params['refit'] = False
optuna_grid_params['n_jobs'] = 2
optuna_grid = OptunaSearchCV(
    estimator=Ridge().set_fit_request(sample_weight=True),
    **optuna_grid_params
)
optuna_grid.fit(X=x, y=y, sample_weight=w, score_param=score_param)


# Fail to observe error when custom metric returns NaN, although error_score='raise'

def custom_metric_with_nan(y: np.ndarray, y_pred: np.ndarray, sample_weight: np.ndarray, score_param: str, *args, **kwargs) -> float:
    return np.nan


optuna_grid_params['scoring'] = make_scorer(
    custom_metric_with_nan,
    greater_is_better=True,
).set_score_request(
    sample_weight=True,
    score_param=True
)
optuna_grid_params['n_jobs'] = 1
optuna_grid = OptunaSearchCV(
    estimator=Ridge().set_fit_request(sample_weight=True),
    **optuna_grid_params
)
optuna_grid.fit(X=x, y=y, sample_weight=w, score_param=score_param)
assert optuna_grid_params['error_score'] == 'raise'

Additional context (optional)

No response