In scikit-learn we keep a clear separation between predictors and transformers.
A predictor’s API is defined around fit/predict, while a transformer is expected to implement fit/transform (and optionally fit_transform)

Predictors do not implement transform because returning predictions is not considered a transformation of the feature space in the scikit-learn design. Doing so would blur the API boundaries and make it harder for tools like Pipeline, FeatureUnion, and meta-estimators to reason about the estimator type. For example, optimizers or wrappers that assume “transformers only change features” would break if a predictor started acting as one.

If you want to feed model predictions as features, the recommended pattern is to wrap the predictor in a thin transformer, e.g:

from sklearn.base import BaseEstimator, TransformerMixin

class PredictorTransformer(BaseEstimator, TransformerMixin):
    def __init__(self, estimator):
        self.estimator = estimator

    def fit(self, X, y=None):
        self.estimator.fit(X, y)
        return self

    def transform(self, X):
        return self.estimator.predict(X).reshape(-1, 1)

Then you can safely use this wrapper inside a FeatureUnion without mixing up the core APIs