This PR currently breaks some tests:

• linear algebra
• sparse arrays
• dates
• TwicePrecision (used for floating-point ranges)

This seems to be quite breaking for somewhat limited gains. A lot of the breakage might happen in the wild, and wouldn't be caught by a PkgEvalrun.

I'm also not fully certain about the motivation here. One of the examples in the issue is iszero(Int) returns false, but probably a specific method should be added for Types to avoid such comparisons? Also, probably a similar method may be added for Ring? After all, iszero(x) is documented to be a check for x == zero(x), where zero(x) is the additive identity. Special cases should generally be added when this isn't what zero returns.

In principle I like this, but we definitely need PkgEval here to evaluate the chaos it could cause.

TwicePrecision (used for floating-point ranges)

This is strictly internal, so doesn't matter. I think this should be broadened to include iszero for arrays since the all-zero array is the additive identity of arrays. But that should really have a specialized definition for efficiency. For isone on matrices you also really want a specialized method. Not sure about Dates... What is the zero date? What is the one date? Seems like this shouldn't really break Dates and maybe that should just be fixed.

Here's one option, very hackey but gets the right answer most of the time:

isone(x::Union{Number, AbstractArray}) = x == one(x)
function isone(x)
    o = one(x)
    if typeof(o) == typeof(x)
        o == x
    else
        throw ArgumentError("...")
    end
end

@nanosoldier runtests()

As far as I remember this still fails a bunch of tests in the test suite here, so running PkgEval is probably just a waste.

I think it only fails on stdlibs so this should be somewhat realistic (though not fully).

The package evaluation job you requested has completed - possible new issues were detected.
The full report is available.

Looking at a few of the nano-soldier results, I estimate this breakage will require fixes to be applied to at least 19 registered packages, even once patches are applied to fix the LinearAlgebra case. This is far too breaking.