The following workflow works

[1: shared = 'data']
data = [1, 2]

[2]
input: for_each = dict(par=data)
print(par)

because shared expose variables to the context of the steps that depends on step 1 in a forward workflow. In contrast, the following workflow

[data: shared = 'data']
data = [1, 2]

[default]
input: for_each = dict(par=data)
print(par)

will not work because default does not depend on data.

Now the tricky part is that

[1: shared = 'data']
data = [1, 2]

[2]
input: '1.txt', for_each = dict(par=data)
print(par)

will also not work because 2 does not depend on 1 on the surface so data is not available when 2 is executed.

To make things clear, I propose that we require, at least document explicit use of sos_variable() for any share. That is to say, we not only use sos_variable to fix the above example,

[1: shared = 'data']
data = [1, 2]

[2]
depends: sos_variable('data')
input: '1.txt', for_each = dict(par=data)
print(par)

but also document the use of sos_variable in the working example:

[1: shared = 'data']
data = [1, 2]

[2]
depends: sos_variable('data')
input: for_each = dict(par=data)
print(par)

In this way at least we do not have to explain why data sometimes exists and sometimes does not exist in examples of this ticket.