Let nanarg{min,max}(..., axis=...) optionally return a placeholder value for all-nan slices.

[anntzer]

Currently, nanarg{min,max}(..., axis=...) raises a ValueError when any full-nan slice exists. On numpy's side it looks like a while ago (#3030) this used to return UINTPTR_MIN (the largest negative value), and that behavior was removed to avoid hiding bugs.

I agree with raising by default, but OTOH this prevents writing code where we know that all-nan slices are possible and we are ready to handle them. Adding an option like raise_on_all_nan: bool (defaulting to True) would allow one to do e.g.

idxs = np.nanargmin(t, axis=1, raise_on_all_nan=False)
all_nans_mask = idxs < 0
# do some handling with all_nans_mask

Similar issue posted on numpy's side as numpy/numpy#12352.

[kwgoodman]

My goal is to follow numpy. So I'd only consider changing if you got them to change.

To follow current numpy behavior I do:

if (allnan) {
        VALUE_ERR("All-NaN slice encountered");
        return NULL;
}