Revert changes to inferred valid age that are not consistent with edges model assumed by precompile

[vtjnash]

Reverts only the code from #58254, not the docs

This adds unsound behavior to the Compiler in a doc PR that is supposed to be about increasing soundness. It isn't easily observable behavior, since the added code only triggers if code gets precompiled and saved and the Method gets used before its official "primary" world, but it can lead to unpredictable and inconsistent inference results.

[vtjnash]

For example, the following code (invoking the method to add 1+1 with assume_effect added) is well-formed (if discouraged), and is not UB as implied by #58254:

julia> w = Base.get_world_counter(); Base.@assume_effects :consistent add(a, b) = a + b; Core.invoke_in_world(w, Core.invoke, add, @which(add(1, 1)), 1, 1)
2

I tried to ask for this code to be removed from the PR, but I guess I wasn't precise enough.

FIXED: nothrow->consistent, since it is only an issue for consistent code (see comment below)

[Keno]

I don't think the revert is productive. Can you please be more precise what your actual concern is?

[Keno]

For example, the following code (invoking the method to add 1+1 with assume_effect added) is well-formed (if discouraged), and is not UB as implied by #58254:

The world age bound affects the :consistent annotation, so the example is not applicable. However, with the :consistent annotation, the point of the change is precisely to not apply the :consistent annotation in this scenario to comply with the tightened specification.

[vtjnash]

Okay, I fixed the example above to use consistent instead. I see: you want to express that the method is not valid before it was defined. That seems still incorrect though, since there's nothing semantically relevant to the method itself about that specific world. The PR also failed to define an edge against that method, so re-validation later will result in this being re-added to the valid range in unpredictable ways. For user-friendliness, we usually also have edges against the opposite branch now too (not required for correctness, but we normally do have them):

else
                update_valid_age!(sv, WorldRange(1, m.primary_world - 1))

update pushed to keep the doc changes

[Keno]

you want to express that the method is not valid before it was defined.

The method is valid, but the consistency annotation is not