As you might know, a number of us are working towards bringing various imputation transformers into Scikit-Learn. In addition to imputation transformers, there are also a number of PRs working to implement direct model estimation in the presence of missing data without explicit imputation as a preprocessing step. In other words, we are slowly moving in the direction of comprehensively supporting preprocessing and analysis based on datasets with missing values. Given this context, I just wanted to start a discussion on whether it would make sense for Scikit Learn to consider making missing-data support either a "requirement" or a "highly recommended" component in all future algorithms that will be added to the library (to the extent that such support makes sense and is established in the literature, of course). In my humble opinion, not only would this approach lead to a more "robust" set of tools but also saves a lot of time and resources compared to refactoring the code at a later time, especially in situations when it is done by somebody other than the original author themselves. I would very much appreciate it if the core developers and other contributors could comment on this issue. Thank you for your time and consideration.

TL;DR: Would it make sense for Scikit Learn to have a default policy of missing value support in all applicable algorithms that are added to the library in the future?