Folded literal factored with its case-sensitive twin: (?i:Z)x|Z matches "z"

[jemmix]

A case-insensitive literal in one alternation arm leaks its fold into the equivalent case-sensitive arm, which is enough for the latter to match lowercase input:

Pattern.compile("(?i:Z)x|Z").matcher("z").find()      // true — no arm should match "z"
                                                       // arm 1 needs an x; arm 2 is case-sensitive Z
Pattern.compile("(?i:[Z])x|[Z]").matcher("z").find()  // true — same via CHAR_CLASS
Pattern.compile("Zx|Z").matcher("z").find()           // false — correct

java.util.regex returns false for the first two.

Cause: alternation factoring (Parser.factor, round 1) merges arms by Regexp.equals(), which for LITERAL/CHAR_CLASS compares only the runes — not FoldCase. A folded literal therefore factors with its case-sensitive twin and the factored prefix carries the fold into arms that never asked for it (in the repro the case-sensitive arm is effectively deleted). Go's regexp/syntax compares Flags&FoldCase in its equality; the port dropped it.

Fix: compare FoldCase in Regexp.equals()/hashCode() for LITERAL/CHAR_CLASS (hash must change too, since factor groups via hashCode first).

Disclosure: found while fuzzing another library (https://github.com/jemmix/tdfa-jvm) against re2j; analysis and patch composed agentically with GLM 5.3 assistance.