[Enfiladeguide]

Fixes #1119 (reproducer in #1120).

Since PHP-Parser 5, an operator decides about its own parentheses from the precedence its parent passes down (pInfixOp(), pPrefixOp(), pPostfixOp()). A node with preserved formatting reuses its original code, so that decision never runs, and a node moved into a context that binds more tightly loses parentheses it needs:

// $x = a() && b();
$stmts[0]->expr->expr = new Expr\BooleanNot($stmts[0]->expr->expr);
// printFormatPreserving(): $x = !a() && b();   expected: $x = !(a() && b());

This PR performs the same check in p() for nodes with preserved formatting:

• preservedOpRequiresParens() is the check of pInfixOp() / pPostfixOp() (against $precedence) and of pPrefixOp() (against $lhsPrecedence);
• a prefix operator at the end of the reused code ($c === $d = $e, $f >> print $g) extends as far to the right as possible, so the trailing operand (getTrailingOperandName()) is printed with the $lhsPrecedence of the new context, as pInfixOp() / pPrefixOp() do, unless the original code already has parentheses around it.

As a result, a moved node gets the same parentheses as a newly created node in the same place, including the existing conservative approximation in pFixup().

New fixture test/code/formatPreservation/fixupMovedNodes.test: the #1119 case, $a ?? ($b ? 'yes' : 'no'), $a - ($b - $c), a move into a preserved parent, trailing prefix operators, and cases that must stay without parentheses (left-associative minus, assignment on the RHS, a slot that already has parentheses). It fails on master and passes here. The full test suite, PHPStan and php-cs-fixer pass (PHP 8.5.4).

To check the change beyond hand-picked cases, we ran a differential fuzz: random expressions are pretty-printed and parsed, one node is wrapped in a new operator, moved to another position, or replaces its parent, and the output of printFormatPreserving() is parsed again and compared with the modified AST. Over 60,000 rewrites (3 seeds):

	wrong meaning	parse error
master	6,235	126
this PR	0	0

832 outputs that were already correct on master change. In 829 of them the output is identical to printing a freshly constructed copy of the moved node in the same slot (the added parentheses come from the lhsPrecedence == precedence approximation in pFixup()); in the other 3 the moved node keeps its own original inner parentheses. Concat (see the FIXME on its precedence) and Pipe were left out of the fuzz. Happy to contribute the fuzz script if it is useful.

We ran into this through Rector's TernaryToNullCoalescingRector, which turned $a !== null ? $a : ($b ? 'yes' : 'no') into $a ?? $b ? 'yes' : 'no' (rectorphp/rector#9912, worked around in the rule in rectorphp/rector-src#8674).

Disclosure: the investigation and this change were prepared with automated tooling; the reproduction, tests and proposed change were checked by a person before posting.