Equivalent types have different behavior with old trait solver

[theemathas]

I tried this code:

pub trait Mirror {
    type Output;
}
impl<T> Mirror for T {
    type Output = T;
}

// Refl<X> = X
pub type Refl<X> = <X as Mirror>::Output;

pub trait Trait {
    type Assoc;
}

fn works<T>()
where
    T: Trait,
    <Refl<T> as Trait>::Assoc: Sized,
{
}

fn fails<T>()
where
    Refl<T>: Trait,
    <Refl<T> as Trait>::Assoc: Sized,
{
}

I expected both works and fails to compile. Instead, fails doesn't compile, producing the following error:

error[E0277]: the trait bound `T: Trait` is not satisfied
  --> src/lib.rs:22:1
   |
22 | / fn fails<T>()
23 | | where
24 | |     Refl<T>: Trait,
25 | |     <Refl<T> as Trait>::Assoc: Sized,
   | |_____________________________________^ the trait `Trait` is not implemented for `T`
   |
help: consider further restricting type parameter `T` with trait `Trait`
   |
25 |     <Refl<T> as Trait>::Assoc: Sized, T: Trait
   |                                       ++++++++

error[E0277]: the trait bound `T: Trait` is not satisfied
  --> src/lib.rs:25:32
   |
25 |     <Refl<T> as Trait>::Assoc: Sized,
   |                                ^^^^^ the trait `Trait` is not implemented for `T`
   |
help: consider further restricting type parameter `T` with trait `Trait`
   |
25 |     <Refl<T> as Trait>::Assoc: Sized, T: Trait
   |                                       ++++++++

For more information about this error, try `rustc --explain E0277`.

Minimized from code written by backslash_phi on the rust community discord

@rustbot labels +A-trait-system +T-types

Meta

Reproducible on the playground with version 1.91.0-nightly (2025-08-10 c8ca44c98eade864824a)

[theemathas]

Another variant of the reproducer: https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=ed5f565abd1752a426e3f6dc32a0c734

[compiler-errors]

This is a duplicate of probably one of the many other "param env normalization" bugs we have. Perhaps scan the fixed-by-next-solver label for one like this?

[theemathas]

Seems similar to #132913