Tim Jenness: Brian Van Klaveren mentioned in Slack that cffi is already pulled into our python environment by other dependencies.  Is it safe for me to just rely on that, or should I create and eups stub package to check that cffi is correctly installed?

At minimum cffi should be added to the list of prereqs that lsstsw has to install from conda (https://github.com/lsst/lsstsw/blob/master/etc/conda3_bleed-linux-64.txt and the macOS file). Please don't rely on it coming from another dependency indirectly. If you do that I'm not overly fussed about a stub eups package in this case.

It looks like, when we ship pybind11 with the stack, we strip it down to just the C++ header files so that running import pybind11 from within python fails.

GalSim version 2.0 uses import pybind11 in its SConstruct file to learn some configuration values (where the pybind11 include files are kept, for instance) during installation.

Is there a reason we strip down pybind11 the way that we do? Am I free to edit our pybind11 repo so that import pybind11 once again works in python?

Jim Bosch will be able to answer your pybind11 question.

For the record: it's actually not actually anything we are doing explicitly. It looks like if you run make install in pybind11, it just copies over the C++ header files. Since we blow away the build directory after installing pybind11, we accidentally drop the python-importable files, which are not copied over to the install directory.

Mike Jarvis is working on a patch to GalSim that make it unnecessary to import pybind11 from within python, so I may not have to touch our pybind11 repo.

It is also possible that we could get around this problem by using python setup.py to install pybind11, instead of make install.

Is there a reason not to do that, Jim Bosch?

Mike has patched GalSim, so the question of how we install pybind11 is no longer immediately pressing.

Ugh, pybind11 appears to be annoying in this respect. I'd be all for getting the the pybind11 Python module installed, but I also want the stuff in $PREFIX/share/cmake installed, and I bet neither of pybind11's two install mechanisms does both. So maybe we should run both?