Treat lookahead/lookbehind as zero-width in RegexGroupParser and look through non-capturing groups in getRootAlternation

[phpstan-bot]

Summary

Regex capturing groups containing lookahead/lookbehind assertions or empty alternation branches inside non-capturing groups were incorrectly inferred as non-empty-string or non-falsy-string when they could match empty strings. This fix teaches RegexGroupParser that lookahead/lookbehind assertions are zero-width (don't consume characters) and that alternation branches should use "any" semantics for maybe-empty checks.

Changes

• src/Type/Regex/RegexGroupParser.php:

  • walkGroupAst: Added early return for #lookahead, #negativelookahead, #lookbehind, #negativelookbehind nodes to prevent their content from influencing nonEmpty/nonFalsy/numeric state
  • isMaybeEmptyNode: Added early return true for lookahead/lookbehind nodes (zero-width assertions). Added special handling for #alternation nodes: an alternation is maybe-empty if ANY branch is maybe-empty (previously used the generic recursion which required ALL children to be maybe-empty). Isolated $isNonFalsy per alternation branch to prevent cross-branch leakage
  • getRootAlternation: Extended to look through #noncapturing groups for both #capturing and #namedcapturing groups, enabling union type creation for patterns like ((?:|\d+)) → ''|numeric-string

• tests/PHPStan/Analyser/nsrt/preg_match_shapes.php:

  • Added bug12840 test function with 9 test cases covering: empty alternation in non-capturing group, lookahead/lookbehind/negative variants in alternation, lone lookahead in group, two non-capturing groups with empty alternation, named capturing groups with non-capturing wrappers, and lookahead in concatenation with literal

Root cause

Three interrelated issues in RegexGroupParser:

1. Lookahead/lookbehind treated as character-consuming: walkGroupAst recursed into their children, causing the content of zero-width assertions (e.g., the x in (?=x)) to set nonEmpty=yes as if the assertion consumed characters. isMaybeEmptyNode similarly didn't recognize them as zero-width.

2. isMaybeEmptyNode used wrong semantics for alternation: The generic recursion checked if ALL children were maybe-empty (correct for concatenation) but alternation only needs ANY branch to be maybe-empty. This meant (?:(?=x)|y) was considered non-empty because the y branch was non-empty, even though the (?=x) branch is zero-width.

3. getRootAlternation didn't see through non-capturing groups: Patterns like ((?:|\d+)) couldn't use the union-type code path (which creates ''|numeric-string) because the alternation was wrapped in a #noncapturing node.

Test

• bug12840 function in tests/PHPStan/Analyser/nsrt/preg_match_shapes.php with 9 assertType calls covering:
  • Empty alternation in non-capturing group: ((?:|\d+)) → ''|numeric-string
  • All four assertion variants in alternation: (?=...), (?<=...), (?!...), (?<!...) → string (not non-falsy-string)
  • Lone lookahead in group: ((?=x)) → string (not non-empty-string)
  • Named group with non-capturing wrapper: (?P<g>(?:|\d+)) → ''|numeric-string
  • Lookahead in concatenation with literal: ((?=x)a) → non-falsy-string (correctly preserved)

Fixes phpstan/phpstan#12840