Negated unicode-set classes are only valid when every member is
single-code-point. We already rejected direct string-valued members
such as `q{ab}` and `p{RGI_Emoji_Flag_Sequence}` inside `[^...]`,
but nested class-set operands could still smuggle them through, so
patterns like `[^[[p{Emoji_Keycap_Sequence}]]]` and the reported
fuzzed literal compiled instead of throwing.
Validate nested class-set expressions after parsing and reject only the
negated `/v` classes whose resulting multi-code-point strings are still
non-empty. Track the exact string members contributed by string
literals, string properties, and nested classes so intersections and
subtractions can eliminate them before the negated-class check runs.
Add constructor and literal coverage for the reduced nested-string
cases, the original regression, and valid negated set operations that
remove every string member.
e0de4ef33e