Improve the Character default mutator: mutate only reaches 8 fixed seeds

[twof]

Summary

Character.defaultMutator's mutate is far weaker than its generate. Mutation can only ever return one of 8 hard-coded seeds, ignoring the input character entirely — so it does no neighborhood or structural exploration and cannot reach most of the character space.

Where

Sources/PropertyTestingKit/Fuzzing/Mutators/MutatorProviding/Character+MutatorProviding.swift

private let _characterSeeds: [Character] = ["a", "Z", "0", " ", "\n", "\t", "😄", "\0"]

private func _characterMutate(_ value: Character, _ rng: inout FastRNG) -> Character {
    // Pick ONE applicable seed (any seed other than the current value).
    let candidates = _characterSeeds.filter { $0 != value }
    guard let mutant = candidates.randomElement(using: &rng) else { return value }
    return mutant
}

Why this is weak

• Tiny reachable set. A mutant is always one of the 8 seeds. From any input, mutation is effectively a 7-way categorical jump; the overwhelming majority of Character values are unreachable as mutants.
• No locality. It never explores near the input — e.g. a → b, case flip a → A, digit 5 → 6, adjacent Unicode scalar. Coverage-guided search benefits from small steps that flip one branch at a time; a categorical jump to a fixed seed can't do that.
• generate already knows more than mutate. _characterGenerate draws from ASCII-printable, lowercase/uppercase letters, digits, whitespace, and an emoji set — a much richer space than the 8 mutation seeds. That structured knowledge isn't used on the mutation path.

Suggested directions

Make mutate pick uniformly among a set of applicable, non-identity strategies (same pattern recently applied to the port/HTTP/percentage/empty-string mutators), e.g.:

• Scalar neighborhood: increment/decrement the input's Unicode scalar by a small delta.
• Case flip: a ↔ A when the input is cased.
• Class-local step: within letters/digits, move to an adjacent member (5 → 4/6).
• Class jump: swap to a random member of a different class (letter ↔ digit ↔ whitespace ↔ symbol ↔ emoji) — reusing the richer pools generate already defines instead of the 8-seed list.
• Keep the existing identity filter so a mutant is never equal to the input.

Acceptance

• mutate reaches a substantially larger and more structured set than the current 8 seeds.
• Mutation includes at least one locality-preserving step (scalar/case/class-local).
• No identity mutants (already covered by IdentityMutantRateTests; add Character-specific coverage asserting reachability of the new strategies).

Context: surfaced while removing identity mutants from the single-value mutators (PR #42).