Fundamentally, I have been thinking that the estimator tags feel more natural attached to the class.

We use the instance when we define "pairwise":

That was a recent addition, where I assumed that tags will stay connected to the instance. (If we want to move tags into the class level, I would greatly prefer "pairwise" to just be an public attribute third party developers can set "X_pairwise_".

Are there any reasons why we have them specifically on the instance?

Meta-estimators I assume, since they must delegate most tags to the inner estimator which is only known after instantiation, and sometimes even after fit()

Meta-estimators I assume, since they must delegate most tags to the inner estimator which is only known after instantiation, and sometimes even after fit()

Hmm, I guess the most visible one is *SearchCV when searching through estimators, *SearchCV should inherit the tags of the estimator that it found. (I do not think this is implemented)

There are other meta-estimators like SFS and RFE, but not as many as I expected:

(sklearn-dev) ➜  scikit-learn git:(main) ✗ git grep -n _safe_tags\(self
sklearn/feature_selection/_base.py:86:            force_all_finite=not _safe_tags(self, key="allow_nan"),
sklearn/feature_selection/_from_model.py:314:        return {"allow_nan": _safe_tags(self.estimator, key="allow_nan")}
sklearn/feature_selection/_rfe.py:415:            "allow_nan": _safe_tags(self.estimator, key="allow_nan"),
sklearn/feature_selection/_sequential.py:236:            "allow_nan": _safe_tags(self.estimator, key="allow_nan"),
sklearn/model_selection/_search.py:382:            "pairwise": _safe_tags(self.estimator, "pairwise"),
sklearn/multiclass.py:577:        return {"pairwise": _safe_tags(self.estimator, key="pairwise")}
sklearn/multiclass.py:880:        return {"pairwise": _safe_tags(self.estimator, key="pairwise")}
sklearn/pipeline.py:642:        return {"pairwise": _safe_tags(self.steps[0][1], "pairwise")}
sklearn/utils/_tags.py:34:    where `est` comes from: typically `_safe_tags(self.base_estimator)` where

(I do not think this is implemented)

Indeed, tag delegation doesn't seem to be fully implemented. But it does seem that we rely on it at least partially, so making the tags a class-level property might be a bit tricky

For good or bad, the parameters of an estimator fundamentally change its capabilities/API at times. I think the answer to this is "sorry, no, find another way".

Ok.