bpart: failed "edge" invalidation

[timholy]

This adds a test for an invalidation failure when consts are redefined. The failure occurs only in "edge" mode, i.e., due to the separation of precompilation from the user's running session.

The corresponding "immediate" sequence is the following (which seems to work properly):

julia> struct InvalidatedBinding
           x::Int
       end

julia> struct Wrapper
           ib::InvalidatedBinding
       end

julia> makewib(x) = Wrapper(InvalidatedBinding(x))
makewib (generic function with 1 method)

julia> const gib = makewib(1)
Wrapper(InvalidatedBinding(1))

julia> fib() = gib.ib.x
fib (generic function with 1 method)

julia> fib()
1

julia> ci = only(methods(fib)).specializations.cache
CodeInstance for MethodInstance for fib()

julia> struct InvalidatedBinding
           x::Float64
       end

julia> ci.max_world
0xffffffffffffffff

julia> struct Wrapper
           ib::InvalidatedBinding
       end

julia> const gib = makewib(2.0)
Wrapper(InvalidatedBinding(2.0))

julia> ci.max_world
0x0000000000009789

julia> fib() === 2.0
true

But it fails when the precompilation of fib happens in a package that gets loaded after gib is redefined in the user's session.

[timholy]

I took it off the milestone and slapped a "merge me" on it, but once merged we should make a milestone item to turn those @test_brokens into @test. EDIT: done in #58346.

[Keno]

tests for this code path are mostly in rebinding.jl. I don't really mind it being here, but I don't fully know yet what distinguishes this case from the ones already tested

[timholy]

Thanks, I hadn't found that yet. Likewise I'd have no objection to moving these tests to rebinding.jl.

[timholy]

One potential difference: from a quick scan, the tests in rebinding.jl seem to define the function in the same module in which it's precompiled. This test defines StaleA.fib() but defers its precompilation until StaleB, and yanks the rug out in between loading those two modules.