Pattern "\uDC21"` matches inside a surrogate pair; equivalent classes don't

[jemmix]

A pattern whose literal prefix is a lone low surrogate matches inside a well-formed surrogate pair; the equivalent alternation/char-class patterns do not:

Pattern.compile("\uDC21").matcher("a\uD801\uDC21b").find()          // true — U+10421's low unit
Pattern.compile("\uDC21|\uDC22").matcher("a\uD801\uDC21b").find()   // false
Pattern.compile("[\uD800-\uDFFF]").matcher("a\uD801\uDC21b").find() // false

Whether a match may start between the units of a pair depends on whether the pattern compiles to a single literal prefix.

Repro (SurrogatePairTest):

assertFalse(Pattern.compile("\uDC21").matcher("a\uD801\uDC21b").find());
// search must resume past the skipped interior hit:
Matcher m = Pattern.compile("\uDC21").matcher("\uD801\uDC21\uDC21");
assertTrue(m.find());
assertEquals(2, m.start());

Cause: MachineInput.StringInput.index() jumps to raw String.indexOf hits. Such a hit can be the low unit of a pair — not a codepoint boundary — and step() then decodes the lone low unit as a rune. Paths without the fast path decode pairs and never enter the interior.

Fix: skip indexOf hits that are a low surrogate preceded by a high surrogate; keep searching.

Sidenote: RE2 proper operates on UTF-8 runes, where pair-interior positions cannot exist; this is a UTF-16 port artifact.

Disclosure: this was discovered when fuzzing another library (https://github.com/jemmix/tdfa-jvm) against re2j. Patch and fix composed agentically using GLM 5.3.