Enumerated the build matrix.

[pelson]

I've just been tinkering with conda-smithy to work as expected for VC versions. I'll be submitting a PR there shortly, but in the meantime, this is what it looks like here.

[jjhelmus]

python needs to be added as a test requirement as a python based test is now included. I'm not exactly sure how the the test were able to pass in the initial render without this, perhaps the system python was being selected?

[pelson]

I'm not exactly sure how the the test were able to pass in the initial render without this

Only one matrix item would have been run, namely CONDA_PY=27. Because that is the version of python being installed into the root environment, the test would presumably have just worked. The problem is that now, we are building all of the VC feature versions, and so the test no longer functions as expected. Adding python to the test requirements should do the trick. I'll iterate shortly.

[pelson]

I've pushed python into the test requirements. Let's see how that does.

[jjhelmus]

I still think that python needs to be a run requirement on Windows to pick up the correct vc feature. Not sure why the Travis CI test is failing, I'll investigate in a few hours at home with my Mac.

[jakirkham]

Sorry to be daft, but is this an attempt to get glpk support for later VC versions?

[pelson]

I still think that python needs to be a run requirement on Windows to pick up the correct vc feature.

If that is the case then we have no need to specify the features ourselves though? We can just rely on python having done that?

[jakirkham]

What if we just set the toolset directly as suggested here?

[jakirkham]

Not that I think we want to follow this exactly, but this spells it out a bit more.

[jakirkham]

In any event, the proposal presented here seems to be doing the right thing for Windows. For example, see the end of this py35 log. It is selecting vc14 correctly.

[jjhelmus]

Sorry for ignoring this the last two days, had some other work that needed attention. I'll see if I can fix the Windows and OS X build issue in another PR.

[jjhelmus]

The strange thing is all the packages were built and passed the tests in conda-forge/staged-recipes#141

[jakirkham]

Basically, I believe this is related to this underlying problem with conda build. Setting the Python versions in the build matrix seems like the sane solution. That way dropping the run requirement on Python should be possible.

[jjhelmus]

Closing in favor of #2. @pelson thanks for looking into this one.