HTMLRestrictions::merge() can crash due to array_merge_recursive() being affected by internal PHP array state

Problem/Motivation

@larowlan reported a crash in #3222797-52: Upgrade path from CKEditor 4's StylesCombo to CKEditor 5's Style, which I could not reproduce by just creating a simple test case. It took many hours of trial and error and debugging.

Eventually, the root cause became clear: only if somewhere along the way while various things are computed and data is transformed, certain arrays have an array state where the "next" array key by appending to the array (so $array[] = 'something';) would not be zero (i.e. $array[0] once existed and was since deleted), then \Drupal\ckeditor5\HTMLRestrictions::merge() computes an invalid result.

But … the root cause is not the logic in HtmlRestrictions::merge(). It's in PHP's own array_merge_recursive()! https://php.net/manual/en/function.array-merge-recursive.php does not document this behavior. The consequence is that two ostensibly identical arrays can yield different merge results! 🤯

$a = ['foo' => ['bar' => TRUE]];
$b = ['foo' => 'baz'];

$m1 = array_merge_recursive($a, $b);

$a['foo'][] = 'proof is in the pudding';
unset($a['foo'][0]);

$m2 = array_merge_recursive($a, $b);

var_dump($m1 == $m2);

👆 This prints bool(false)! 🤯🤯🤯

Steps to reproduce

See https://3v4l.org/fXXsn, where I came up with a way to demonstrate the problem and show that there's only once possible solution:

<?php

$a = ['p' => ['id' => TRUE]];
$b = ['p' => FALSE];

print "Behaves as expected.\n";
var_dump(array_merge_recursive($a, $b));

print "\n\nOMG internal array pointer shenanigans can affect the results of array_merge_recursive()!\n";
$a['p'][0] = 'SOMETHING, anything!';
unset($a['p'][0]);
var_dump(array_merge_recursive($a, $b));

print "\n\nThis keeps happening…\n";
var_dump(array_merge_recursive($a, $b));

print "\n\n… not even resetting the internal array pointer helps …\n";
reset($a['p']);
var_dump(array_merge_recursive($a, $b));

print "\n\n… nor casting to object, cloning, then back to array …\n";
$a['p'] = (array) clone (object) $a['p'];
var_dump(array_merge_recursive($a, $b));


print "\n\n… only forcefully recreating a copy-by-value of the array by using array_slice().\n";
$a['p'] = array_slice($a['p'], 0);
var_dump(array_merge_recursive($a, $b));

outputs:

Behaves as expected.
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [0]=>
    bool(false)
  }
}


OMG internal array pointer shenanigans can affect the results of array_merge_recursive()!
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [1]=>
    bool(false)
  }
}


This keeps happening…
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [1]=>
    bool(false)
  }
}


… not even resetting the internal array pointer helps …
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [1]=>
    bool(false)
  }
}


… nor casting to object, cloning, then back to array …
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [1]=>
    bool(false)
  }
}


… only forcefully recreating a copy-by-value of the array by using array_slice().
array(1) {
  ["p"]=>
  array(2) {
    ["id"]=>
    bool(true)
    [0]=>
    bool(false)
  }
}

Proposed resolution

Work around this weakness in array_merge_recursive() by ensuring that the "next" array key of every nested array is always zero.

Remaining tasks

1. Tests
2. Fix
3. Reviews

User interface changes

None.=

API changes

None.

Data model changes

None.

Release notes snippet

None.