Incompleteness in preferring subst-relate candidate in alias-relate goals

[compiler-errors]

There are 3 ways to prove AliasRelate goals: lhs normalize-to rhs, rhs normalize-to lhs, and substs-relate.

If merging the three options via fn try_merge_responses fails, we incompletely prefer substs-relate over normalizing. https://github.com/rust-lang/rust/blob/a482149598f5aacf3837eee87026dd634f08641c/compiler/rustc_trait_selection/src/solve/alias_relate.rs#L84-L86

It seems like this incompleteness is not necessary if we go with #25 instead.

---

edit: original description

ill write this tomorrow since it's 21:50 rn and i want to sleep

tl;dr here's the list of failing tests before and after -- red (left-hand side) tests are those that are fixed by the incompleteness, and green (right-hand side) are those that begin to fail after the incompleteness: https://www.diffchecker.com/NdEJkqfr/

[compiler-errors]

if we were to change this incompleteness to apply only if the response has no infer/opaques, this is the list of failing tests for each scenario: https://www.diffchecker.com/PbEqtAfC/ -- given that these two incompletenesses only differ by ~10 tests or so, it suggests most of the test fail -> pass come from preferring subst-relate only when regions are involved.

this suggests that we may be able to get away (for now) with only preferring a subst-relate candidate if it has_only_region_constraints.

[lcnr]

This is used to fix #2

[compiler-errors]

We don't do this anymore as of rust-lang/rust#113901 -- bidirectional normalizes-to is both more conservative and is sufficient in most cases, and #2 was also fixed rust-lang/rust#112399.