Currently, data with nan's are either always propagated, or an error is raised (e.g. np.asarray_chkfinite raises a ValueError).
The behaviour around nan's is currently unspecified and untested.

Luckily, SciPy has their very well-specified nan_policy (https://scipy.github.io/devdocs/dev/api-dev/nan_policy.html) which appears to be very much applicable to (at least) Lmo's sample estimators.