For a tuple primal @scalar_rule returns a tuple (not a Composite) differential.

[oxinabox]

With #163 we effectily decided that as far as we are concerned, julia only has single output functions.
in #163 we made the approprate changes for rrules which was that pullbacks need to take a single input, which might be a Composite{Tuple}.

But we have not made the change for forward mode, at least not consistently or to the @scalar_rule macro.

An example (and possibly the only example) of this problem is:

We currently have

julia> frule((Zero(), 1.0), sincos, π)
((1.2246467991473532e-16, -1.0), (-1.0, -1.2246467991473532e-16))

But we should have:

julia> frule((Zero(), 1.0), sincos, π)
((1.2246467991473532e-16, -1.0), Composite{Tuple{Float64, Float64}}(-1.0, -1.2246467991473532e-16))

This should be changed in this section:

ChainRulesCore.jl/src/rule_definition_tools.jl

Lines 157 to 162 in 0fe9da8

	if n_outputs > 1
	# For forward-mode we only return a tuple if output actually a tuple.
	pushforward_returns = Expr(:tuple, pushforward_returns...)
	else
	pushforward_returns = pushforward_returns[1]
	end

This issiue becomes more pressing with #207 because that macro can't trivially know if a function is returnig multiple outputs, but if we say multiple outputs don't exist, then it is easy.