Evaluate caches stack-dependent results

[lcnr]

struct W<T: ?Sized>(*const T);
trait Foo {}
trait Bar {}

impl<T: Bar + NotImplemented> Foo for T {}
impl Foo for W<u32> {}

impl<T: Foo + AssertInferenceConstraintsApplied> Bar for T {}
trait AssertInferenceConstraintsApplied {}
trait GuideFromEnv {}
trait ErrOnGuidance {}
impl<T: GuideFromEnv + ErrOnGuidance> AssertInferenceConstraintsApplied for T {}
impl<T> GuideFromEnv for T {}
impl ErrOnGuidance for W<u32> {}

impl<T> Bar for T
where
    W<T>: NotImplemented {}
trait NotImplemented {}

fn impls_foo<T: Foo>() {}
fn impls_bar<T: Bar>() {}

fn with_bound()
where
    W<u64>: GuideFromEnv,
{
    impls_foo::<W<_>>(); // commenting this line changes the next one to OK
    impls_bar::<W<_>>(); // ERROR
}

fn main() {
    with_bound();
}

cc @compiler-errors #123303 this should be broken even after #122791

THe idea is from #123303 (comment)

• GoalA
  • CandA1
    • GoalB
      • CandB1
        • GoalA (inductive cycle -> ambig)
        • AssertInferenceConstraintsApplied (fails due to incompleteness if the inference constraints from CandA2 have not yet been applied)
      • CandB2
        • Impossible (necessary to evaluate nested goals of CandB1 during selection)
  • CandA2
    • guide inference ~> Ok, but inference constraints