`inference_triggers` says root is child node

[luke-kiernan]

When I try to run inference_triggers, I get the following:

julia> itrigs = inference_triggers(tinf)
ERROR: ArgumentError: it seems you've supplied a child node, but backtraces are collected only at the entrance to inference.

You can even see it in the docs. Is it that the docs are out-of-date and there's some other method I should be using? Or is this a legitimate bug?

[manuelbb-upb]

Same problem after updating Julia from 1.11.4 to 1.12.5.

[manuelbb-upb]

I have for now downgraded to 1.11 again;
Could this issue be related to the changes in JuliaLang/julia#53415?
Adapting my example from JuliaDebug/Cthulhu.jl#693 I can provoke the problem:

using SnoopCompileCore

function top(
    args... 
)
    if isempty(args)
        return nothing
    end
    return sub(args...)
end

function sub(
    args...
)
    return top(args[1:end-1]...)
end

tinf = @snoop_inference top(1, 2.0)
for child in tinf.children
    @show child.ci
    @show isnothing(child.bt)
end

Gives

child.ci = CodeInstance for MethodInstance for top(::Int64, ::Vararg{Any})
isnothing(child.bt) = false
child.ci = CodeInstance for MethodInstance for sub(::Vararg{Any})
isnothing(child.bt) = true
child.ci = CodeInstance for MethodInstance for top()
isnothing(child.bt) = false

I have tried investigating further by extracting the macro code for @snoop_inference and looking at the code instances, but my knowledge about compiler internals is too limited.
@timholy In timingtree we are caching on method instances instead of code instances to reverse the CodeInstance edges. Could we instead not directly use the backedges field of MethodInstances?
In the example above, tinf.children[2].def.backedges is defined.
However, if I unpack the macro, I get more children, some of which do not have backedges for their method instance defined. Is this the result of invalidation?