Checking 'did it warn before removal?': call the real path with real values

When you check whether a removed keyword warned in the previous release, constructing the object with the keyword is not enough. In scikit-learn 1.7.2/1.8.0, ColumnTransformer(force_int_remainder_cols=...) is silent at construction and emits the FutureWarning only in fit(). Warnings can also depend on the value. statsmodels 0.14.5 AutoReg(old_names=True) warns, while old_names=False (the default) is silent. Test whether the keyword worked at all before the removal, too. In statsmodels 0.14.5, RLM.fit(init=<any truthy value>) already raised AttributeError, and acorr_lm(cov_kwargs=...) never reached the covariance estimator (HAC with cov_kwargs={'maxlags': 2} raises KeyError). Their removal in 0.15 only changes the error type. Call construct + fit/predict under warnings.catch_warnings(record=True) with simplefilter('always'), for both a truthy and a falsy value, before classifying a removal as 'breaks working code without warning'.