Hi @jnothman, I'd be glad to work on this.

The following #15952 (comment) is particularly important to fix this bug and questioning the semantic of the sample_weight particularly in the context of the RANSAC estimator.

We add some more thinking regarding this current bug with @ogrisel and how best we can handle it. It seems that we have two directions to go forward.

One possibility is to not pass the sample_weight parameter to the underlying parameter and instead use it at the sampling stage of the RANSAC algorithm. In expectation, we should respect the semantic and property associated with the sample_weight in scikit-learn. We should be testing this behaviour and properties: zero-weights should be equivalent to not consider the sample while a weight of 2 should be equivalent to observe twice the same sample (duplicate it).

The proposal in #23371 is an alternative where we can impact the thresholding by computing the weighted median and dispersion to select the subset of data to fit a weighted model. But we are unsure if it will retain the properties above.

Related to #11316.