Michael Abbott

Yes to #31. For now this is explicitly defined to just fall through: https://github.com/mcabbott/AxisKeys.jl/blob/f822f269b0dd4271521b8518561fc8507f6d9e06/src/functions.jl#L386-L391

This case should work with https://github.com/JuliaDiff/ChainRules.jl/pull/655, but Diffractor's own rule takes precedence.

This is tricky. My first thought here is that, without thinking about this package, you probably want to just reshape and call `*` instead of making slices. One package which...

What's the argument against defining these? Besides being one less thing to know about, `@test x ≈ 0.0` will I think give more useful information when it fails, as the...

Accumulating N terms naiively needs 2N-1 allocations: N to make the terms and then N-1 to add them pairwise. If each term makes a thunk, this PR would reduce that...

Xref discussion at #581, I guess that it would be slightly better for this to return a ZeroTangent than a NoTangent.

The stronger way to abstract this would just be to make the Tangent constructor return a Zero. I don't quite see why an all-Zero Tangent is ever better.

Good point. Ideally as many as possible of these would run all ways. The extent of my attempt at this so far is [these lines](https://github.com/JuliaDiff/Diffractor.jl/pull/73/files#diff-fba4d4a7c64aacd9b1a662bca4487f3e849b43f09e65f6c8788a3f142b989f65R20-R28) , which ran into https://github.com/JuliaDiff/Diffractor.jl/issues/70...

These tests really want a different `@test_broken` macro, which accepts an error as a broken test, but fails on `false` -- a silently wrong answer is much worse than an...

Yes I think that should work, as should `a[down=1] .+= b[down=1]`. But they don't, both are bugs I think.