math.atan2(0, 0) (or likewise with floats) returns 0 or pi depending on the signedness of the zeros in Py2 and Py3 and that's also mandated by the C standard. norm_text_angle later takes it modulo 180 so we're fine.

I see where you're coming from but mixing in a ternary in a if... elif chain looks meh to me.

Actually, I don't think that the three blocks are logically equal. To me, the current form

if self._sort_zpos is not None:
    [use _sort_zpos]
elif tzs.size > 0:
    [use min(tzs)]
else:
    [use fallback]

is logically more like

if self._sort_zpos is not None:
    [use _sort_zpos]
else:
    if tzs.size > 0:
        [use min(tzs)]
    else:
        [use fallback]

Usually tzs are expected to be there and nan is only a fallback if they are not. Thus it would be preferable to

if self._sort_zpos is not None:
    [use _sort_zpos]
else:
    [use min(tzs) with fallback]

However, I agree that depends on the point of view you take. And I won't argue further if you still prefer to leave it as is.

I know (hence the "I see where you're coming from" above), but it still looks not so nice to me. (Really, best would be min(tzs, default=np.nan) but np.min doesn't have that and builtins.min is likely slower.)

Slightly better:

default_z = 0  # FIXME: arbitrary default
zsortval = np.min(zvals) if zvals.size else default_z

went for some middle ground

Can this be indented to align with the bracket in the line above?

not really, that'll look awkward in the actual docstring (the problem is the same as always re: line continuations in rst).

I think treating the first line of the docstring as a replacement __signature__ is a sphinx thing, not an rst thing - it's possible it gets handled before the RST parser kicks in.