there should be a way to tag the estimator with conditions like "does only take 1d input". Currently this is hard-coded for Isotonic in the tests.
This should be configurable exactly for your use case.
There is currently an estimator_type attribute that does something related (declare classifiers and regressors) but that's not powerful enough. We probably need to add a classifier_tags property or something like that (though how do you set that through inheritance? have it be a property that calls the parent property?)

Actually, given that isotonic regression has a natural multidimensional generalization (#5744), perhaps it should be instead changed to take a 2d-input.
Still, I agree that a marker "takes only 1D input" could be generally useful (for example I certainly don't plan to write a multidimensional generalization to near-isotonic regression).

see #6599.