However, keep using 4.9.1 on the non-latest builds. Though we do support
back to 4.9.0, that version is not available in conda.

This is a known bug fix in proj.4 5.0.0, wherein lat_0=0 used to
erroneously multiply results by 2.

See OSGeo/PROJ#194

If using exactly central_longitude+/-180, one boundary will now wrap
around and produce the position of the opposite one. So subtract a small
epsilon from that boundary to prevent wrap-around. This produces
something on the tenth of a millimetre difference, but that probably
won't be visible when plotted.

Seems like those releases were a bit buggy, since it goes back to the
previous result in 5.1.0.

That is, any comments, documentation, or output, now drops the '4'.
Names in code have not been changed to preserve backwards-compatibility.

It does not work correctly in 5.0.0 <= proj4 < 5.1.0.

The Google Mercator projection seems to come out slightly smaller than
it should.