ReferenceConstraintViolation with non-empty-literal-string in array

[AndrolGenhald]

non-empty-literal-string should be a subtype of non-empty-string, but it still results in a ReferenceConstraintViolation: https://psalm.dev/r/2756501f64

This seems specific to ReferenceConstraintViolation, it acts correctly when a non-empty-literal-string is used as an argument that requires non-empty-string: https://psalm.dev/r/f8845a750d

In fact, it seems to be something about using it in an array, just having it as the parameter itself works fine: https://psalm.dev/r/c9ad19355c

[psalm-github-bot]

I found these snippets:

https://psalm.dev/r/2756501f64

<?php

/**
 * @param "a"|"b"|"c" $a
 * @param array<non-empty-string, true> $b
 */
function foo(string $a, array &$b): void
{
    /** @psalm-trace $baz */
    $baz = "bar $a";
    $b[$baz] = true;
}

$b = [];
foo("a", $b);

Psalm output (using commit 0ca856b):

INFO: Trace - 10:5 - $baz: non-empty-literal-string

ERROR: ReferenceConstraintViolation - 7:32 - Variable $b is limited to values of type array<non-empty-string, true> because it is passed by reference, non-empty-array<string, true> type found. Use @param-out to specify a different output type

https://psalm.dev/r/f8845a750d

<?php

/**
 * @param non-empty-string $bar
 */
function foo(string $bar): string
{
    return $bar;
}

/** @var "bar"|"baz" */
$bar = "bar";
/** @psalm-trace $baz */
$baz = "foo$bar";
foo($baz);

Psalm output (using commit 0ca856b):

INFO: Trace - 14:1 - $baz: non-empty-literal-string

https://psalm.dev/r/c9ad19355c

<?php

/**
 * @param "a"|"b"|"c" $a
 * @param non-empty-string $b
 */
function foo(string $a, string &$b): void
{
    /** @psalm-trace $baz */
    $baz = "bar $a";
    $b = $baz;
}

$b = "b";
foo("a", $b);

Psalm output (using commit 0ca856b):

INFO: Trace - 10:5 - $baz: non-empty-literal-string

[craigfrancis]

I'm not familiar with Psalm internals (and I've run out of time today), but it looks like $all_types_contain is getting set to false in ArrayTypeComparator.php (I assume that shouldn't be the case).