`obligations_for_self_ty` leaks placeholders

[lcnr]

trait Foo<'a> {}
fn needs_foo<T>(_: T)
where
    for<'a> Wrap<T>: Foo<'a>,
{}

struct Wrap<T>(T);
impl<'a, T> Foo<'a> for Wrap<T> where T: Fn(&'a i32) {}

fn main() {
    needs_foo(|x| println!("{x}"));
}

During typeck, the signature of |x| println!("{x}") gets inferred to fn(&'!a i32) where '!a is a placeholder which is not nameable by the closure. This is unsound as in "it breaks a core invariant of the type system". As all regions are erased after hir typeck and rechecked during mir borrowck it should not cause any "allows for UB at runtime" unsoundness.

This is still something we should fix, ideally by "pulling down" the inference variables of the closure signature into the root universe when deducing it. The same issue exists for all uses of obligations_for_self_ty.

This may cause some ICEs in the future if we improve our validation of type system invariants in the future.

cc @rust-lang/types

edit: for now this is actually not too different from how higher ranked relate works in general. However, the same issue exists with feature(non_lifetime_binders) and we want to more eagerly handle these placeholder constraints in the future to handle where-bounds on binders

[RodBurman]

The latest stable toolchain:

% cargo -v -V
cargo 1.85.0 (d73d2caf9 2024-12-31)
release: 1.85.0
commit-hash: d73d2caf9e41a39daf2a8d6ce60ec80bf354d2a7
commit-date: 2024-12-31
host: aarch64-apple-darwin
libgit2: 1.8.1 (sys:0.19.0 vendored)
libcurl: 8.7.1 (sys:0.4.74+curl-8.9.0 system ssl:(SecureTransport) LibreSSL/3.3.6)
ssl: OpenSSL 1.1.1w  11 Sep 2023
os: Mac OS 15.3.2 [64-bit]

Fails to build this code with the following error:

% cargo build
   Compiling o4st_leaks v0.1.0 (/Users/rod/code/rust/triage/o4st_leaks)
error: implementation of `Foo` is not general enough
  --> src/main.rs:11:5
   |
11 |     needs_foo(|x| println!("{x}"));
   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ implementation of `Foo` is not general enough
   |
   = note: `Wrap<{closure@src/main.rs:11:15: 11:18}>` must implement `Foo<'0>`, for any lifetime `'0`...
   = note: ...but it actually implements `Foo<'1>`, for some specific lifetime `'1`

error: could not compile `o4st_leaks` (bin "o4st_leaks") due to 1 previous error

I am not sure that was the desired outcome for the issue, but it should (may?) prevent any future ICE.