TypeError: in typeassert, expected Union{Nothing, Base.ScopedValues.Scope}

[lgoettgens]

Since yesterday we see the following error in our nightly CI (see thofma/Hecke.jl#1687), and something similar locally

TypeError: in typeassert, expected Union{Nothing, Base.ScopedValues.Scope}, got a value of type ZZRingElem
  Stacktrace:
    [1] get
      @ ./scopedvalues.jl:152 [inlined]
    [2] _precision_with_base_2
      @ ./mpfr.jl:1041 [inlined]
    [3] BigFloat
      @ ./mpfr.jl:150 [inlined]
[...]

  TypeError: in typeassert, expected Union{Nothing, Base.ScopedValues.Scope}, got a value of type Tuple{Base.MP
FR.MPFRRoundingMode}
  Stacktrace:
    [1] get
      @ ./scopedvalues.jl:152 [inlined]
    [2] rounding_raw
      @ ./mpfr.jl:241 [inlined]
    [3] BigFloat
      @ ./mpfr.jl:369 [inlined]
[...]

From the set of changes since this appeared I assume this is due to #55907.

Originally posted by @lgoettgens in #55907 (comment)

[thofma]

Or sometimes

 Got exception outside of a @test
 TypeError: in typeassert, expected Union{Nothing, Base.ScopedValues.Scope}, got a value of type Core.Const
 Stacktrace:
   [1] setprecision(f::Nemo.var"#1023#1024"{ArbFieldElem}, ::Type{BigFloat}, prec::Int64; base::Int64)
     @ Base.MPFR ./mpfr.jl:1180
   [2] setprecision
     @ ./mpfr.jl:1179 [inlined]
   [3] round
     @ ~/.julia/packages/Nemo/0TzOv/src/HeckeMoreStuff.jl:157 [inlined]

I am sure I did not create a Core.Const somewhere myself on purpose.

[topolarity]

The fix in #56612 appears to be working (Hecke.jl tests pass for me locally now), but I am quite confused how a Core.Const was making it into Core.current_scope()...

We were forgetting to restore the current scope, but we were never putting anything that's not a Base.ScopedValues.Scope into ct->scope so I'm confused how we produced the strange values in the failures above.

[topolarity]

I am quite confused how a Core.Const was making it into Core.current_scope()...

I think I figured this out.

The problem is that we were using the scope value after it was dead from the point-of-view of our GC passes, which is illegal. By chance we were observing other objects that overlapped with us in memory. That made it look like non-Scope values were making it into ct->scope, but they were not.