Places where we still use isscalar internally sensibly:

• assert_equal - probably harmless
• linspace - used as an optimization check without breaking duck-types

Places where we shouldn't be using it:

• np.poly1d - Only used in an optimization path. Since this is pseudo-deprecated anyway, it's not worth touching this..
• np.matrix - Cause of a bug - m[:, np.array(0)] is the transpose of a[:, 0]. Not worth fixing since matrix is in many ways broken by design
• np.piecewise - used for an undocumented special case that probably shouldn't exist in the first place.
• np.ma.apply_along_axis - Cause of a bug, fixed in MAINT: make np.ma.apply_along_axis consistent with np.apply_along_axis #8511

Places where we no longer do because it caused issues:

• np.gradient (gradient() no longer accepts scalar-like wrappers for dx #8292)
• np.histogram (MAINT: Tidy np.histogram, and improve error messages #9889)
• np.histogramdd (BUG/DOC/MAINT: Tidy up histogramdd #10802)
• np.apply_along_axis (BUG: Fix crash on 0d return value in apply_along_axis #8441)

Very informative--I learned a lot just reading this PR.

@charris, can you restart the appveyor build?

Although probably safe to ignore, since this is a doc-only change

Apparently it is not possible to restart single jobs on AppVeyor

Right now it's not even possible for me to restart the whole appveyor run, since the appveyor account is not the org one (as discussed elsewhere).

Seems wasteful to rerun all the tests on a doc-only just because appveyor failed

Seems wasteful to rerun all the tests on a doc-only just because appveyor failed

agreed

Thanks Eric.