fugue-evo 0.3.0

//! Mutation operators
//!
//! This module provides various mutation operators for genetic algorithms.

use rand::Rng;
use rand_distr::{Distribution, Normal};

use crate::genome::bit_string::BitString;
use crate::genome::bounds::MultiBounds;
use crate::genome::permutation::Permutation;
use crate::genome::real_vector::RealVector;
use crate::genome::traits::{
    BinaryGenome, EvolutionaryGenome, PermutationGenome, RealValuedGenome,
};
use crate::genome::tree::{Function, Terminal, TreeGenome, TreeNode};
use crate::operators::traits::{BoundedMutationOperator, MutationOperator};

/// Polynomial mutation (bounded)
///
/// Uses the polynomial probability distribution to perturb genes.
/// Respects bounds and is commonly used with NSGA-II.
///
/// Reference: Deb, K. (2001). Multi-Objective Optimization using Evolutionary Algorithms.
#[derive(Clone, Debug)]
pub struct PolynomialMutation {
    /// Distribution index (typically 20-100)
    /// Higher values = smaller mutations
    pub eta_m: f64,
    /// Per-gene mutation probability (default: 1/n)
    pub mutation_probability: Option<f64>,
    /// Standard deviation used by the *unbounded* fallback (see
    /// [`MutationOperator::mutate`]). `None` selects the adaptive default
    /// `0.1 * (1 + |x|)` per gene; `Some(s)` uses a fixed `s`.
    pub unbounded_sigma: Option<f64>,
}

impl PolynomialMutation {
    /// Create a new polynomial mutation with the given distribution index
    pub fn new(eta_m: f64) -> Self {
        assert!(eta_m >= 0.0, "Distribution index must be non-negative");
        Self {
            eta_m,
            mutation_probability: None,
            unbounded_sigma: None,
        }
    }

    /// Set a fixed mutation probability per gene
    pub fn with_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.mutation_probability = Some(probability);
        self
    }

    /// Set the standard deviation of the unbounded Gaussian fallback.
    ///
    /// Polynomial mutation is only defined relative to finite bounds; when the
    /// operator is invoked without bounds it perturbs each gene with Gaussian
    /// noise instead (audit EV-102). By default the per-gene sigma is
    /// `0.1 * (1 + |x|)`; this method pins it to a fixed value.
    pub fn with_unbounded_sigma(mut self, sigma: f64) -> Self {
        assert!(sigma >= 0.0, "Sigma must be non-negative");
        self.unbounded_sigma = Some(sigma);
        self
    }

    /// Apply polynomial mutation to a gene
    fn mutate_gene<R: Rng>(&self, gene: f64, min: f64, max: f64, rng: &mut R) -> f64 {
        let range = max - min;
        if range <= 0.0 {
            return gene;
        }

        let delta1 = (gene - min) / range;
        let delta2 = (max - gene) / range;

        let u = rng.gen::<f64>();
        let delta_q = if u <= 0.5 {
            let val = 2.0 * u + (1.0 - 2.0 * u) * (1.0 - delta1).powf(self.eta_m + 1.0);
            val.powf(1.0 / (self.eta_m + 1.0)) - 1.0
        } else {
            let val = 2.0 * (1.0 - u) + 2.0 * (u - 0.5) * (1.0 - delta2).powf(self.eta_m + 1.0);
            1.0 - val.powf(1.0 / (self.eta_m + 1.0))
        };

        (gene + delta_q * range).clamp(min, max)
    }
}

impl MutationOperator<RealVector> for PolynomialMutation {
    fn mutate<R: Rng>(&self, genome: &mut RealVector, rng: &mut R) {
        // Polynomial mutation is intrinsically bounds-relative and is undefined
        // without a finite range. Rather than fabricating +/-1e10 bounds (which
        // turned an "unbounded" mutation into a destructive, near-random reset
        // of each gene), fall back to a local Gaussian perturbation (EV-102).
        let n = genome.dimension();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);

        for gene in genome.genes_mut() {
            if rng.gen::<f64>() < prob {
                let sigma = self
                    .unbounded_sigma
                    .unwrap_or_else(|| 0.1 * (1.0 + gene.abs()));
                if sigma > 0.0 {
                    let normal = Normal::new(0.0, sigma).unwrap();
                    *gene += normal.sample(rng);
                }
            }
        }
    }

    fn mutation_probability(&self) -> Option<f64> {
        self.mutation_probability
    }
}

impl BoundedMutationOperator<RealVector> for PolynomialMutation {
    fn mutate_bounded<R: Rng>(&self, genome: &mut RealVector, bounds: &MultiBounds, rng: &mut R) {
        let n = genome.dimension();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);

        for i in 0..n {
            if rng.gen::<f64>() < prob {
                if let Some(bound) = bounds.get(i) {
                    genome.genes_mut()[i] =
                        self.mutate_gene(genome.genes()[i], bound.min, bound.max, rng);
                }
            }
        }
    }
}

/// Gaussian mutation
///
/// Adds Gaussian noise to each gene.
#[derive(Clone, Debug)]
pub struct GaussianMutation {
    /// Standard deviation of the Gaussian noise
    pub sigma: f64,
    /// Per-gene mutation probability
    pub mutation_probability: Option<f64>,
}

impl GaussianMutation {
    /// Create a new Gaussian mutation with the given standard deviation
    pub fn new(sigma: f64) -> Self {
        assert!(sigma >= 0.0, "Sigma must be non-negative");
        Self {
            sigma,
            mutation_probability: None,
        }
    }

    /// Set a fixed mutation probability per gene
    pub fn with_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.mutation_probability = Some(probability);
        self
    }
}

impl MutationOperator<RealVector> for GaussianMutation {
    fn mutate<R: Rng>(&self, genome: &mut RealVector, rng: &mut R) {
        let n = genome.dimension();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);
        let normal = Normal::new(0.0, self.sigma).unwrap();

        for gene in genome.genes_mut() {
            if rng.gen::<f64>() < prob {
                *gene += normal.sample(rng);
            }
        }
    }

    fn mutation_probability(&self) -> Option<f64> {
        self.mutation_probability
    }
}

impl BoundedMutationOperator<RealVector> for GaussianMutation {
    fn mutate_bounded<R: Rng>(&self, genome: &mut RealVector, bounds: &MultiBounds, rng: &mut R) {
        let n = genome.dimension();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);
        let normal = Normal::new(0.0, self.sigma).unwrap();

        for i in 0..n {
            if rng.gen::<f64>() < prob {
                genome.genes_mut()[i] += normal.sample(rng);
                if let Some(bound) = bounds.get(i) {
                    genome.genes_mut()[i] = bound.clamp(genome.genes()[i]);
                }
            }
        }
    }
}

/// Uniform mutation
///
/// Replaces genes with random values within bounds.
#[derive(Clone, Debug)]
pub struct UniformMutation {
    /// Per-gene mutation probability
    pub mutation_probability: Option<f64>,
}

impl UniformMutation {
    /// Create a new uniform mutation
    pub fn new() -> Self {
        Self {
            mutation_probability: None,
        }
    }

    /// Set a fixed mutation probability per gene
    pub fn with_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.mutation_probability = Some(probability);
        self
    }
}

impl Default for UniformMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<RealVector> for UniformMutation {
    fn mutate<R: Rng>(&self, genome: &mut RealVector, rng: &mut R) {
        // Uniform mutation requires a finite range; without explicit bounds we
        // fall back to the modest, non-destructive default range [-1, 1].
        // Prefer [`mutate_bounded`] with real bounds whenever they are known.
        let default_bounds = MultiBounds::symmetric(1.0, genome.dimension());
        self.mutate_bounded(genome, &default_bounds, rng);
    }

    fn mutation_probability(&self) -> Option<f64> {
        self.mutation_probability
    }
}

impl BoundedMutationOperator<RealVector> for UniformMutation {
    fn mutate_bounded<R: Rng>(&self, genome: &mut RealVector, bounds: &MultiBounds, rng: &mut R) {
        let n = genome.dimension();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);

        for i in 0..n {
            if rng.gen::<f64>() < prob {
                if let Some(bound) = bounds.get(i) {
                    genome.genes_mut()[i] = rng.gen_range(bound.min..=bound.max);
                }
            }
        }
    }
}

/// Bit-flip mutation for bit strings
///
/// Flips each bit with a given probability.
#[derive(Clone, Debug)]
pub struct BitFlipMutation {
    /// Per-bit mutation probability (default: 1/n)
    pub mutation_probability: Option<f64>,
}

impl BitFlipMutation {
    /// Create a new bit-flip mutation
    pub fn new() -> Self {
        Self {
            mutation_probability: None,
        }
    }

    /// Set a fixed mutation probability per bit
    pub fn with_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.mutation_probability = Some(probability);
        self
    }
}

impl Default for BitFlipMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<BitString> for BitFlipMutation {
    fn mutate<R: Rng>(&self, genome: &mut BitString, rng: &mut R) {
        let n = genome.len();
        let prob = self.mutation_probability.unwrap_or(1.0 / n as f64);

        for i in 0..n {
            if rng.gen::<f64>() < prob {
                genome.flip(i);
            }
        }
    }

    fn mutation_probability(&self) -> Option<f64> {
        self.mutation_probability
    }
}

/// Swap mutation for permutation genomes (also works on any genome)
///
/// Swaps two random positions in the genome.
#[derive(Clone, Debug)]
pub struct SwapMutation {
    /// Number of swaps to perform
    pub num_swaps: usize,
}

impl SwapMutation {
    /// Create a new swap mutation with a single swap
    pub fn new() -> Self {
        Self { num_swaps: 1 }
    }

    /// Create with multiple swaps
    pub fn with_swaps(num_swaps: usize) -> Self {
        Self { num_swaps }
    }
}

impl Default for SwapMutation {
    /// Delegates to [`SwapMutation::new`] (a single swap). A derived `Default`
    /// would set `num_swaps = 0`, producing a silent no-op operator (EV-101).
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<BitString> for SwapMutation {
    fn mutate<R: Rng>(&self, genome: &mut BitString, rng: &mut R) {
        let n = genome.len();
        if n < 2 {
            return;
        }

        for _ in 0..self.num_swaps {
            let i = rng.gen_range(0..n);
            let j = rng.gen_range(0..n);
            if i != j {
                let temp = genome.bits()[i];
                genome.bits_mut()[i] = genome.bits()[j];
                genome.bits_mut()[j] = temp;
            }
        }
    }
}

impl MutationOperator<RealVector> for SwapMutation {
    fn mutate<R: Rng>(&self, genome: &mut RealVector, rng: &mut R) {
        let n = genome.dimension();
        if n < 2 {
            return;
        }

        for _ in 0..self.num_swaps {
            let i = rng.gen_range(0..n);
            let j = rng.gen_range(0..n);
            if i != j {
                genome.genes_mut().swap(i, j);
            }
        }
    }
}

/// Scramble mutation
///
/// Scrambles a random segment of the genome.
#[derive(Clone, Debug, Default)]
pub struct ScrambleMutation;

impl ScrambleMutation {
    /// Create a new scramble mutation
    pub fn new() -> Self {
        Self
    }
}

impl MutationOperator<BitString> for ScrambleMutation {
    fn mutate<R: Rng>(&self, genome: &mut BitString, rng: &mut R) {
        use rand::seq::SliceRandom;

        let n = genome.len();
        if n < 2 {
            return;
        }

        let mut start = rng.gen_range(0..n);
        let mut end = rng.gen_range(0..n);
        if start > end {
            std::mem::swap(&mut start, &mut end);
        }

        // Extract segment, shuffle, and put back
        let segment: Vec<bool> = (start..=end).map(|i| genome.bits()[i]).collect();
        let mut shuffled = segment;
        shuffled.shuffle(rng);

        for (i, val) in shuffled.into_iter().enumerate() {
            genome.bits_mut()[start + i] = val;
        }
    }
}

impl MutationOperator<RealVector> for ScrambleMutation {
    fn mutate<R: Rng>(&self, genome: &mut RealVector, rng: &mut R) {
        use rand::seq::SliceRandom;

        let n = genome.dimension();
        if n < 2 {
            return;
        }

        let mut start = rng.gen_range(0..n);
        let mut end = rng.gen_range(0..n);
        if start > end {
            std::mem::swap(&mut start, &mut end);
        }

        // Shuffle the segment in place
        let slice = &mut genome.genes_mut()[start..=end];
        slice.shuffle(rng);
    }
}

// =============================================================================
// Permutation Mutation Operators
// =============================================================================

/// Swap mutation for permutation genomes
///
/// Swaps two random positions in the permutation.
/// This is one of the simplest and most commonly used permutation mutations.
#[derive(Clone, Debug)]
pub struct PermutationSwapMutation {
    /// Number of swaps to perform
    pub num_swaps: usize,
}

impl PermutationSwapMutation {
    /// Create a new swap mutation with a single swap
    pub fn new() -> Self {
        Self { num_swaps: 1 }
    }

    /// Create with multiple swaps
    pub fn with_swaps(num_swaps: usize) -> Self {
        Self { num_swaps }
    }
}

impl Default for PermutationSwapMutation {
    /// Delegates to [`PermutationSwapMutation::new`] (a single swap). A derived
    /// `Default` would set `num_swaps = 0`, a silent no-op operator (EV-101).
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<Permutation> for PermutationSwapMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        let n = genome.dimension();
        if n < 2 {
            return;
        }

        for _ in 0..self.num_swaps {
            let i = rng.gen_range(0..n);
            let j = rng.gen_range(0..n);
            if i != j {
                genome.swap(i, j);
            }
        }
    }
}

/// Insert mutation for permutation genomes
///
/// Removes an element from one position and inserts it at another.
/// This preserves adjacencies better than swap mutation.
#[derive(Clone, Debug)]
pub struct InsertMutation {
    /// Number of insert operations to perform
    pub num_inserts: usize,
}

impl InsertMutation {
    /// Create a new insert mutation with a single insert
    pub fn new() -> Self {
        Self { num_inserts: 1 }
    }

    /// Create with multiple inserts
    pub fn with_inserts(num_inserts: usize) -> Self {
        Self { num_inserts }
    }
}

impl Default for InsertMutation {
    /// Delegates to [`InsertMutation::new`] (a single insert). A derived
    /// `Default` would set `num_inserts = 0`, a silent no-op operator (EV-101).
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<Permutation> for InsertMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        let n = genome.dimension();
        if n < 2 {
            return;
        }

        for _ in 0..self.num_inserts {
            let from = rng.gen_range(0..n);
            let to = rng.gen_range(0..n);
            if from != to {
                genome.insert(from, to);
            }
        }
    }
}

/// Inversion mutation (2-opt) for permutation genomes
///
/// Reverses a random segment of the permutation.
/// This is particularly effective for TSP-like problems as it can
/// remove crossing edges.
#[derive(Clone, Debug, Default)]
pub struct InversionMutation;

impl InversionMutation {
    /// Create a new inversion mutation
    pub fn new() -> Self {
        Self
    }
}

impl MutationOperator<Permutation> for InversionMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        let n = genome.dimension();
        if n < 2 {
            return;
        }

        let mut start = rng.gen_range(0..n);
        let mut end = rng.gen_range(0..n);
        if start > end {
            std::mem::swap(&mut start, &mut end);
        }

        genome.reverse_segment(start, end);
    }
}

/// Scramble mutation for permutation genomes
///
/// Shuffles a random segment of the permutation.
/// More disruptive than inversion, but still preserves some structure.
#[derive(Clone, Debug, Default)]
pub struct PermutationScrambleMutation;

impl PermutationScrambleMutation {
    /// Create a new scramble mutation
    pub fn new() -> Self {
        Self
    }
}

impl MutationOperator<Permutation> for PermutationScrambleMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        use rand::seq::SliceRandom;

        let n = genome.dimension();
        if n < 2 {
            return;
        }

        let mut start = rng.gen_range(0..n);
        let mut end = rng.gen_range(0..n);
        if start > end {
            std::mem::swap(&mut start, &mut end);
        }

        // Shuffle the segment
        let perm = genome.permutation_mut();
        perm[start..=end].shuffle(rng);
    }
}

/// Displacement mutation for permutation genomes
///
/// Selects a segment, removes it, and inserts it at a random position.
/// This is similar to insert mutation but operates on segments.
#[derive(Clone, Debug, Default)]
pub struct DisplacementMutation;

impl DisplacementMutation {
    /// Create a new displacement mutation
    pub fn new() -> Self {
        Self
    }
}

impl MutationOperator<Permutation> for DisplacementMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        let n = genome.dimension();
        if n < 3 {
            return;
        }

        // Select segment
        let mut start = rng.gen_range(0..n);
        let mut end = rng.gen_range(0..n);
        if start > end {
            std::mem::swap(&mut start, &mut end);
        }

        let segment_len = end - start + 1;
        if segment_len >= n {
            return; // Can't displace the entire permutation
        }

        // Extract segment
        let perm = genome.permutation_mut();
        let segment: Vec<usize> = perm[start..=end].to_vec();

        // Remove segment
        let remaining: Vec<usize> = perm[..start]
            .iter()
            .chain(perm[end + 1..].iter())
            .copied()
            .collect();

        // Choose insertion point in remaining
        let insert_pos = rng.gen_range(0..=remaining.len());

        // Rebuild permutation
        let new_perm: Vec<usize> = remaining[..insert_pos]
            .iter()
            .chain(segment.iter())
            .chain(remaining[insert_pos..].iter())
            .copied()
            .collect();

        perm.copy_from_slice(&new_perm);
    }
}

/// Adaptive mutation rate for permutation genomes
///
/// Combines multiple mutation operators with configurable probabilities.
#[derive(Clone, Debug)]
pub struct AdaptivePermutationMutation {
    /// Probability of swap mutation
    pub swap_prob: f64,
    /// Probability of insert mutation
    pub insert_prob: f64,
    /// Probability of inversion mutation
    pub inversion_prob: f64,
    /// Probability of scramble mutation
    pub scramble_prob: f64,
}

impl AdaptivePermutationMutation {
    /// Create with default probabilities (each equally likely)
    pub fn new() -> Self {
        Self {
            swap_prob: 0.25,
            insert_prob: 0.25,
            inversion_prob: 0.25,
            scramble_prob: 0.25,
        }
    }

    /// Create with custom probabilities
    pub fn with_probs(swap: f64, insert: f64, inversion: f64, scramble: f64) -> Self {
        Self {
            swap_prob: swap,
            insert_prob: insert,
            inversion_prob: inversion,
            scramble_prob: scramble,
        }
    }
}

impl Default for AdaptivePermutationMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl MutationOperator<Permutation> for AdaptivePermutationMutation {
    fn mutate<R: Rng>(&self, genome: &mut Permutation, rng: &mut R) {
        let total = self.swap_prob + self.insert_prob + self.inversion_prob + self.scramble_prob;
        if total <= 0.0 {
            return;
        }

        let r = rng.gen::<f64>() * total;
        let mut cumulative = 0.0;

        cumulative += self.swap_prob;
        if r < cumulative {
            PermutationSwapMutation::new().mutate(genome, rng);
            return;
        }

        cumulative += self.insert_prob;
        if r < cumulative {
            InsertMutation::new().mutate(genome, rng);
            return;
        }

        cumulative += self.inversion_prob;
        if r < cumulative {
            InversionMutation::new().mutate(genome, rng);
            return;
        }

        PermutationScrambleMutation::new().mutate(genome, rng);
    }
}

// =============================================================================
// Tree (GP) Mutation Operators
// =============================================================================

/// Point mutation for tree genomes (genetic programming)
///
/// Selects a random node and replaces it with a new random node of the same
/// type. For function nodes, the replacement has the same arity. For terminal
/// nodes, another random terminal is selected.
///
/// This is a non-destructive mutation that preserves tree structure while
/// changing individual nodes.
#[derive(Clone, Debug)]
pub struct PointMutation {
    /// Per-node mutation probability
    pub mutation_probability: f64,
    /// Probability of selecting a function node (vs terminal)
    pub function_probability: f64,
}

impl PointMutation {
    /// Create a new point mutation with default settings
    ///
    /// Defaults: 0.1 per-node probability, 0.9 function probability
    pub fn new() -> Self {
        Self {
            mutation_probability: 0.1,
            function_probability: 0.9,
        }
    }

    /// Set the per-node mutation probability
    pub fn with_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.mutation_probability = probability;
        self
    }

    /// Set the function selection probability
    pub fn with_function_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.function_probability = probability;
        self
    }

    /// Mutate a single node *in place*, preserving its arity.
    ///
    /// A terminal is replaced with a fresh random terminal; a function is swapped
    /// for a randomly chosen function of the same arity (its children are left
    /// untouched, so tree structure is preserved). In-place mutation via `&mut`
    /// (rather than returning a new owned node) is what lets the whole traversal
    /// avoid moving values out of an owned `TreeNode`, which would conflict with
    /// `TreeNode`'s stack-safe `Drop` impl (EV-60).
    fn mutate_node_in_place<T: Terminal, F: Function, R: Rng>(
        &self,
        node: &mut TreeNode<T, F>,
        rng: &mut R,
    ) {
        match node {
            TreeNode::Terminal(t) => *t = T::random(rng),
            TreeNode::Function(func, _children) => {
                let target_arity = func.arity();
                let matching_funcs: Vec<&F> = F::functions()
                    .iter()
                    .filter(|f| f.arity() == target_arity)
                    .collect();
                if !matching_funcs.is_empty() {
                    *func = matching_funcs[rng.gen_range(0..matching_funcs.len())].clone();
                }
                // Children are preserved: point mutation keeps arity/structure.
            }
        }
    }
}

impl Default for PointMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl<T: Terminal, F: Function> MutationOperator<TreeGenome<T, F>> for PointMutation {
    fn mutate<R: Rng>(&self, genome: &mut TreeGenome<T, F>, rng: &mut R) {
        // Iterative (explicit-stack) in-place point mutation (EV-60): visit every
        // node without recursion so a pathologically deep tree cannot overflow the
        // call stack. Each node is independently mutated with probability
        // `mutation_probability`; point mutation preserves arity, so children are
        // left in place and only the node's own label may change.
        let mut stack: Vec<&mut TreeNode<T, F>> = vec![&mut genome.root];
        while let Some(node) = stack.pop() {
            if rng.gen::<f64>() < self.mutation_probability {
                self.mutate_node_in_place(node, rng);
            }
            if let TreeNode::Function(_, children) = node {
                for child in children.iter_mut() {
                    stack.push(child);
                }
            }
        }
    }

    fn mutation_probability(&self) -> Option<f64> {
        Some(self.mutation_probability)
    }
}

/// Subtree mutation for tree genomes (genetic programming)
///
/// Replaces a randomly selected subtree with a new randomly generated subtree.
/// This is a more disruptive mutation than point mutation.
#[derive(Clone, Debug)]
pub struct SubtreeMutation {
    /// Maximum depth of the generated subtree
    pub max_subtree_depth: usize,
    /// Probability of selecting a function node for replacement
    pub function_probability: f64,
    /// Terminal probability for grow method
    pub terminal_probability: f64,
}

impl SubtreeMutation {
    /// Create a new subtree mutation with default settings
    pub fn new() -> Self {
        Self {
            max_subtree_depth: 4,
            function_probability: 0.9,
            terminal_probability: 0.3,
        }
    }

    /// Set the maximum depth for generated subtrees
    pub fn with_max_depth(mut self, depth: usize) -> Self {
        self.max_subtree_depth = depth;
        self
    }

    /// Set the function selection probability
    pub fn with_function_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.function_probability = probability;
        self
    }

    /// Set the terminal probability for tree generation
    pub fn with_terminal_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.terminal_probability = probability;
        self
    }
}

impl Default for SubtreeMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl<T: Terminal, F: Function> MutationOperator<TreeGenome<T, F>> for SubtreeMutation {
    fn mutate<R: Rng>(&self, genome: &mut TreeGenome<T, F>, rng: &mut R) {
        // Decide whether to select a function or terminal position
        let position = if rng.gen::<f64>() < self.function_probability {
            genome
                .random_function_position(rng)
                .unwrap_or_else(|| genome.random_terminal_position(rng).unwrap_or_default())
        } else {
            genome
                .random_terminal_position(rng)
                .unwrap_or_else(|| genome.random_function_position(rng).unwrap_or_default())
        };

        // Depth budget for the replacement subtree (EV-27).
        //
        // `position` has one entry per edge from the root, so the selected node
        // sits at tree level `position.len() + 1` (the root is level 1).
        // Replacing it with a subtree `S` yields a tree whose depth along that
        // branch is `position.len() + S.depth()`. To keep the whole tree within
        // `genome.max_depth` we require `S.depth() <= max_depth - position.len()`.
        let point_depth = position.len();
        let budget = genome
            .max_depth
            .saturating_sub(point_depth)
            .min(self.max_subtree_depth)
            .max(1);

        // `TreeGenome::generate_grow(_, m, _)` can produce a subtree whose
        // `depth()` is up to `m + 1` (a function node may be created at the last
        // permitted level and still receives terminal children one level
        // deeper). We therefore ask for `budget - 1` so the result's depth is at
        // most `budget`.
        let new_root =
            TreeGenome::<T, F>::generate_grow(rng, budget - 1, self.terminal_probability).root;

        // Defensive guard mirroring SubtreeCrossover: if the generated subtree
        // would still violate the depth limit, fall back to a single terminal,
        // which always fits since `budget >= 1`.
        let new_root = if point_depth + new_root.depth() > genome.max_depth {
            TreeNode::Terminal(T::random(rng))
        } else {
            new_root
        };

        // Replace the subtree
        genome.root.replace_subtree(&position, new_root);
    }
}

/// Hoist mutation for tree genomes (genetic programming)
///
/// Selects a random subtree and replaces the entire tree with it.
/// This is useful for bloat control.
#[derive(Clone, Debug, Default)]
pub struct HoistMutation {
    /// Probability of selecting a function node
    pub function_probability: f64,
}

impl HoistMutation {
    /// Create a new hoist mutation
    pub fn new() -> Self {
        Self {
            function_probability: 0.5,
        }
    }

    /// Set the function selection probability
    pub fn with_function_probability(mut self, probability: f64) -> Self {
        assert!(
            (0.0..=1.0).contains(&probability),
            "Probability must be in [0, 1]"
        );
        self.function_probability = probability;
        self
    }
}

impl<T: Terminal, F: Function> MutationOperator<TreeGenome<T, F>> for HoistMutation {
    fn mutate<R: Rng>(&self, genome: &mut TreeGenome<T, F>, rng: &mut R) {
        // Select a random position (prefer function nodes to make it interesting)
        let position = if rng.gen::<f64>() < self.function_probability {
            genome
                .random_function_position(rng)
                .unwrap_or_else(|| genome.random_terminal_position(rng).unwrap_or_default())
        } else {
            genome.random_position(rng)
        };

        // Skip if selecting the root (no change)
        if position.is_empty() {
            return;
        }

        // Get the subtree and make it the new root
        if let Some(subtree) = genome.root.get_subtree(&position) {
            genome.root = subtree.clone();
        }
    }
}

/// Shrink mutation for tree genomes (genetic programming)
///
/// Replaces a randomly selected subtree with one of its terminals.
/// This reduces tree size and helps with bloat control.
#[derive(Clone, Debug, Default)]
pub struct ShrinkMutation;

impl ShrinkMutation {
    /// Create a new shrink mutation
    pub fn new() -> Self {
        Self
    }
}

impl<T: Terminal, F: Function> MutationOperator<TreeGenome<T, F>> for ShrinkMutation {
    fn mutate<R: Rng>(&self, genome: &mut TreeGenome<T, F>, rng: &mut R) {
        // Find a function node to shrink
        if let Some(func_position) = genome.random_function_position(rng) {
            // Skip root to maintain some structure
            if func_position.is_empty() {
                return;
            }

            // Get a terminal from within the selected subtree
            if let Some(subtree) = genome.root.get_subtree(&func_position) {
                let terminal_positions = subtree.terminal_positions();
                if !terminal_positions.is_empty() {
                    // Pick a random terminal from the subtree
                    let term_pos = &terminal_positions[rng.gen_range(0..terminal_positions.len())];

                    // Get the terminal value
                    if let Some(terminal_node) = subtree.get_subtree(term_pos) {
                        let replacement = terminal_node.clone();
                        // Replace the function node with the terminal
                        genome.root.replace_subtree(&func_position, replacement);
                    }
                }
            }
        }
    }
}

/// Adaptive mutation for tree genomes (genetic programming)
///
/// Combines multiple tree mutations with configurable probabilities.
#[derive(Clone, Debug)]
pub struct AdaptiveTreeMutation {
    /// Probability of point mutation
    pub point_prob: f64,
    /// Probability of subtree mutation
    pub subtree_prob: f64,
    /// Probability of hoist mutation
    pub hoist_prob: f64,
    /// Probability of shrink mutation
    pub shrink_prob: f64,
}

impl AdaptiveTreeMutation {
    /// Create with default probabilities
    pub fn new() -> Self {
        Self {
            point_prob: 0.4,
            subtree_prob: 0.3,
            hoist_prob: 0.15,
            shrink_prob: 0.15,
        }
    }

    /// Create with custom probabilities
    pub fn with_probs(point: f64, subtree: f64, hoist: f64, shrink: f64) -> Self {
        Self {
            point_prob: point,
            subtree_prob: subtree,
            hoist_prob: hoist,
            shrink_prob: shrink,
        }
    }
}

impl Default for AdaptiveTreeMutation {
    fn default() -> Self {
        Self::new()
    }
}

impl<T: Terminal, F: Function> MutationOperator<TreeGenome<T, F>> for AdaptiveTreeMutation {
    fn mutate<R: Rng>(&self, genome: &mut TreeGenome<T, F>, rng: &mut R) {
        let total = self.point_prob + self.subtree_prob + self.hoist_prob + self.shrink_prob;
        if total <= 0.0 {
            return;
        }

        let r = rng.gen::<f64>() * total;
        let mut cumulative = 0.0;

        cumulative += self.point_prob;
        if r < cumulative {
            PointMutation::new().mutate(genome, rng);
            return;
        }

        cumulative += self.subtree_prob;
        if r < cumulative {
            SubtreeMutation::new().mutate(genome, rng);
            return;
        }

        cumulative += self.hoist_prob;
        if r < cumulative {
            HoistMutation::new().mutate(genome, rng);
            return;
        }

        ShrinkMutation::new().mutate(genome, rng);
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use approx::assert_relative_eq;

    #[test]
    fn test_polynomial_mutation_respects_bounds() {
        let mut rng = rand::thread_rng();
        let bounds = MultiBounds::symmetric(5.0, 10);

        for _ in 0..100 {
            let mut genome = RealVector::generate(&mut rng, &bounds);
            let mutation = PolynomialMutation::new(20.0).with_probability(1.0);
            mutation.mutate_bounded(&mut genome, &bounds, &mut rng);

            for (i, &gene) in genome.genes().iter().enumerate() {
                let bound = bounds.get(i).unwrap();
                assert!(
                    gene >= bound.min && gene <= bound.max,
                    "Gene {} out of bounds: {} not in [{}, {}]",
                    i,
                    gene,
                    bound.min,
                    bound.max
                );
            }
        }
    }

    #[test]
    fn test_polynomial_mutation_changes_genome() {
        let mut rng = rand::thread_rng();
        let bounds = MultiBounds::symmetric(5.0, 10);
        let original = RealVector::zeros(10);
        let mut genome = original.clone();

        let mutation = PolynomialMutation::new(20.0).with_probability(1.0);
        mutation.mutate_bounded(&mut genome, &bounds, &mut rng);

        // At least some genes should have changed
        let changed = genome
            .genes()
            .iter()
            .zip(original.genes())
            .filter(|(&a, &b)| a != b)
            .count();
        assert!(changed > 0, "No genes were mutated");
    }

    #[test]
    fn test_polynomial_mutation_eta_effect() {
        let mut rng = rand::thread_rng();
        let bounds = MultiBounds::symmetric(1.0, 1);

        // Low eta = larger mutations
        let low_eta = PolynomialMutation::new(1.0).with_probability(1.0);
        // High eta = smaller mutations
        let high_eta = PolynomialMutation::new(100.0).with_probability(1.0);

        let mut low_total_change = 0.0;
        let mut high_total_change = 0.0;
        let trials = 1000;

        for _ in 0..trials {
            let mut genome_low = RealVector::new(vec![0.0]);
            let mut genome_high = RealVector::new(vec![0.0]);

            low_eta.mutate_bounded(&mut genome_low, &bounds, &mut rng);
            high_eta.mutate_bounded(&mut genome_high, &bounds, &mut rng);

            low_total_change += genome_low[0].abs();
            high_total_change += genome_high[0].abs();
        }

        assert!(
            low_total_change > high_total_change,
            "Low eta should produce larger average changes"
        );
    }

    #[test]
    fn test_gaussian_mutation_changes_genome() {
        let mut rng = rand::thread_rng();
        let original = RealVector::zeros(10);
        let mut genome = original.clone();

        let mutation = GaussianMutation::new(0.1).with_probability(1.0);
        mutation.mutate(&mut genome, &mut rng);

        assert_ne!(genome, original);
    }

    #[test]
    fn test_gaussian_mutation_bounded() {
        let mut rng = rand::thread_rng();
        let bounds = MultiBounds::symmetric(1.0, 10);

        for _ in 0..100 {
            let mut genome = RealVector::zeros(10);
            let mutation = GaussianMutation::new(10.0).with_probability(1.0);
            mutation.mutate_bounded(&mut genome, &bounds, &mut rng);

            for (i, &gene) in genome.genes().iter().enumerate() {
                let bound = bounds.get(i).unwrap();
                assert!(gene >= bound.min && gene <= bound.max);
            }
        }
    }

    #[test]
    fn test_uniform_mutation() {
        let mut rng = rand::thread_rng();
        let bounds = MultiBounds::symmetric(1.0, 10);

        for _ in 0..100 {
            let mut genome = RealVector::zeros(10);
            let mutation = UniformMutation::new().with_probability(1.0);
            mutation.mutate_bounded(&mut genome, &bounds, &mut rng);

            for (i, &gene) in genome.genes().iter().enumerate() {
                let bound = bounds.get(i).unwrap();
                assert!(gene >= bound.min && gene <= bound.max);
            }
        }
    }

    #[test]
    fn test_bit_flip_mutation() {
        let mut rng = rand::thread_rng();
        let original = BitString::zeros(100);
        let mut genome = original.clone();

        let mutation = BitFlipMutation::new().with_probability(0.5);
        mutation.mutate(&mut genome, &mut rng);

        // About half should be flipped
        let flipped = genome.count_ones();
        assert!(
            flipped > 20 && flipped < 80,
            "Expected ~50 flips, got {}",
            flipped
        );
    }

    #[test]
    fn test_bit_flip_mutation_default_probability() {
        let mut rng = rand::thread_rng();
        let original = BitString::zeros(100);
        let mut genome = original.clone();

        let mutation = BitFlipMutation::new(); // 1/n probability
        mutation.mutate(&mut genome, &mut rng);

        // With 1/100 probability, expect ~1 flip on average
        // But due to randomness, we just check some change occurred
        // over multiple trials
        let mut total_flips = 0;
        for _ in 0..100 {
            let mut g = BitString::zeros(100);
            mutation.mutate(&mut g, &mut rng);
            total_flips += g.count_ones();
        }

        // Average should be close to 1
        let avg = total_flips as f64 / 100.0;
        assert!(avg > 0.5 && avg < 2.0, "Expected avg ~1, got {}", avg);
    }

    #[test]
    fn test_swap_mutation() {
        let mut rng = rand::thread_rng();
        let mut genome = RealVector::new(vec![0.0, 1.0, 2.0, 3.0, 4.0]);

        let mutation = SwapMutation::new();
        mutation.mutate(&mut genome, &mut rng);

        // The sum should be preserved
        let sum: f64 = genome.genes().iter().sum();
        assert_relative_eq!(sum, 10.0);
    }

    #[test]
    fn test_swap_mutation_multiple() {
        let mut rng = rand::thread_rng();
        let original: Vec<f64> = (0..10).map(|i| i as f64).collect();
        let mut genome = RealVector::new(original.clone());

        let mutation = SwapMutation::with_swaps(5);
        mutation.mutate(&mut genome, &mut rng);

        // Sum should be preserved
        let sum: f64 = genome.genes().iter().sum();
        assert_relative_eq!(sum, 45.0);
    }

    #[test]
    fn test_scramble_mutation() {
        let mut rng = rand::thread_rng();
        let original: Vec<f64> = (0..10).map(|i| i as f64).collect();
        let mut genome = RealVector::new(original.clone());

        let mutation = ScrambleMutation::new();
        mutation.mutate(&mut genome, &mut rng);

        // Sum should be preserved
        let sum: f64 = genome.genes().iter().sum();
        assert_relative_eq!(sum, 45.0);

        // All values should still be present
        let mut sorted: Vec<f64> = genome.genes().to_vec();
        sorted.sort_by(|a, b| a.partial_cmp(b).unwrap());
        assert_eq!(sorted, original);
    }

    #[test]
    fn test_scramble_mutation_bitstring() {
        let mut rng = rand::thread_rng();
        let original = BitString::new(vec![true, true, true, false, false, false, true, false]);
        let mut genome = original.clone();

        let mutation = ScrambleMutation::new();
        mutation.mutate(&mut genome, &mut rng);

        // Count should be preserved
        assert_eq!(genome.count_ones(), original.count_ones());
    }

    // =========================================================================
    // Permutation Mutation Tests
    // =========================================================================

    #[test]
    fn test_permutation_swap_mutation() {
        let mut rng = rand::thread_rng();
        let original = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);
        let mut genome = original.clone();

        let mutation = PermutationSwapMutation::new();
        mutation.mutate(&mut genome, &mut rng);

        // Should still be valid permutation
        assert!(genome.is_valid_permutation());
        assert_eq!(genome.dimension(), 8);
    }

    #[test]
    fn test_permutation_swap_mutation_multiple() {
        let mut rng = rand::thread_rng();
        let original = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7, 8, 9]);
        let mut genome = original.clone();

        let mutation = PermutationSwapMutation::with_swaps(5);
        mutation.mutate(&mut genome, &mut rng);

        // Should still be valid permutation
        assert!(genome.is_valid_permutation());
        assert_eq!(genome.dimension(), 10);
    }

    #[test]
    fn test_insert_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..50 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);

            let mutation = InsertMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
            assert_eq!(genome.dimension(), 8);
        }
    }

    #[test]
    fn test_inversion_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..50 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);

            let mutation = InversionMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
            assert_eq!(genome.dimension(), 8);
        }
    }

    #[test]
    fn test_inversion_mutation_reverses_segment() {
        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(42);

        let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);

        let mutation = InversionMutation::new();
        mutation.mutate(&mut genome, &mut rng);

        // Should still be valid permutation
        assert!(genome.is_valid_permutation());
    }

    #[test]
    fn test_permutation_scramble_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..50 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);

            let mutation = PermutationScrambleMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
            assert_eq!(genome.dimension(), 8);
        }
    }

    #[test]
    fn test_displacement_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..50 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7, 8, 9]);

            let mutation = DisplacementMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
            assert_eq!(genome.dimension(), 10);
        }
    }

    #[test]
    fn test_adaptive_permutation_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..100 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);

            let mutation = AdaptivePermutationMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
            assert_eq!(genome.dimension(), 8);
        }
    }

    #[test]
    fn test_adaptive_permutation_mutation_custom_probs() {
        let mut rng = rand::thread_rng();

        // Test with only inversion mutation
        let mutation = AdaptivePermutationMutation::with_probs(0.0, 0.0, 1.0, 0.0);

        for _ in 0..50 {
            let mut genome = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);
            mutation.mutate(&mut genome, &mut rng);

            assert!(genome.is_valid_permutation());
        }
    }

    // =========================================================================
    // Tree (GP) Mutation Tests
    // =========================================================================

    use crate::genome::tree::{ArithmeticFunction, ArithmeticTerminal};

    fn create_test_tree() -> TreeGenome<ArithmeticTerminal, ArithmeticFunction> {
        // Create: (+ x0 (* 1.0 x1))
        let x0 = TreeNode::terminal(ArithmeticTerminal::Variable(0));
        let c1 = TreeNode::terminal(ArithmeticTerminal::Constant(1.0));
        let x1 = TreeNode::terminal(ArithmeticTerminal::Variable(1));
        let mul = TreeNode::function(ArithmeticFunction::Mul, vec![c1, x1]);
        let add = TreeNode::function(ArithmeticFunction::Add, vec![x0, mul]);
        TreeGenome::new(add, 5)
    }

    #[test]
    fn test_point_mutation_preserves_structure() {
        let mut rng = rand::thread_rng();
        let original = create_test_tree();
        let original_size = original.size();

        for _ in 0..50 {
            let mut genome = original.clone();
            let mutation = PointMutation::new().with_probability(1.0);
            mutation.mutate(&mut genome, &mut rng);

            // Point mutation preserves tree structure (size)
            assert_eq!(genome.size(), original_size);
            assert!(genome.evaluate(&[1.0, 2.0]).is_finite());
        }
    }

    #[test]
    fn test_point_mutation_changes_tree() {
        let mut rng = rand::thread_rng();
        let original = create_test_tree();
        let mut any_changed = false;

        for _ in 0..100 {
            let mut genome = original.clone();
            let mutation = PointMutation::new().with_probability(1.0);
            mutation.mutate(&mut genome, &mut rng);

            // Check if the evaluation changed (indicates mutation occurred)
            let orig_val = original.evaluate(&[1.0, 2.0]);
            let new_val = genome.evaluate(&[1.0, 2.0]);
            if (orig_val - new_val).abs() > 1e-10 {
                any_changed = true;
                break;
            }
        }

        assert!(
            any_changed,
            "Point mutation should sometimes change the tree"
        );
    }

    #[test]
    fn test_point_mutation_deep_tree_no_stack_overflow() {
        // regression: EV-60 — PointMutation must traverse iteratively so a
        // ~100k-deep tree can be mutated without overflowing the call stack. The
        // previous recursive `mutate_recursive` overflowed at this depth.
        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(7);
        let depth = 100_000usize;
        let mut root: TreeNode<ArithmeticTerminal, ArithmeticFunction> =
            TreeNode::terminal(ArithmeticTerminal::Constant(1.0));
        for _ in 0..depth {
            root = TreeNode::function(ArithmeticFunction::Neg, vec![root]);
        }
        let mut genome = TreeGenome::new(root, depth + 1);
        let size_before = genome.size();

        // mutate every node (probability 1.0) — must not overflow.
        PointMutation::new()
            .with_probability(1.0)
            .mutate(&mut genome, &mut rng);

        // Point mutation preserves structure (arity/size) even at extreme depth.
        assert_eq!(genome.size(), size_before);
        // Free iteratively so the test itself doesn't overflow on teardown.
        genome.dismantle();
    }

    #[test]
    fn test_subtree_mutation() {
        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(42);
        let original = create_test_tree();

        for _ in 0..50 {
            let mut genome = original.clone();
            let mutation = SubtreeMutation::new().with_max_depth(3);
            mutation.mutate(&mut genome, &mut rng);

            // Tree should still be valid and have at least one node
            assert!(genome.size() >= 1);
            // Subtree mutation may generate expressions that are NaN/Inf for some inputs
            // (e.g., division by zero), so we only check structural validity
            let result = genome.evaluate(&[1.0, 2.0]);
            assert!(result.is_nan() || result.is_finite());
        }
    }

    #[test]
    fn test_hoist_mutation_reduces_tree() {
        let mut rng = rand::thread_rng();

        // Create a deeper tree
        let tree: TreeGenome<ArithmeticTerminal, ArithmeticFunction> =
            TreeGenome::generate_full(&mut rng, 4, 5);

        let original_size = tree.size();

        for _ in 0..50 {
            let mut genome = tree.clone();
            let mutation = HoistMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            // Hoist mutation should reduce or maintain tree size
            assert!(genome.size() <= original_size);
            assert!(genome.evaluate(&[1.0, 2.0]).is_finite());
        }
    }

    #[test]
    fn test_shrink_mutation_reduces_tree() {
        let mut rng = rand::thread_rng();

        // Create a deeper tree
        let tree: TreeGenome<ArithmeticTerminal, ArithmeticFunction> =
            TreeGenome::generate_full(&mut rng, 4, 5);

        for _ in 0..50 {
            let mut genome = tree.clone();
            let original_size = genome.size();
            let mutation = ShrinkMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            // Shrink mutation should reduce or maintain tree size
            assert!(genome.size() <= original_size);
            assert!(genome.evaluate(&[1.0, 2.0]).is_finite());
        }
    }

    #[test]
    fn test_adaptive_tree_mutation() {
        let mut rng = rand::thread_rng();

        for _ in 0..100 {
            let mut genome: TreeGenome<ArithmeticTerminal, ArithmeticFunction> =
                TreeGenome::generate_ramped_half_and_half(&mut rng, 2, 5);

            let mutation = AdaptiveTreeMutation::new();
            mutation.mutate(&mut genome, &mut rng);

            // Tree should still be valid
            assert!(genome.size() >= 1);
            assert!(genome.evaluate(&[1.0, 2.0]).is_finite());
        }
    }

    #[test]
    fn test_adaptive_tree_mutation_custom_probs() {
        let mut rng = rand::thread_rng();

        // Test with only point mutation
        let mutation = AdaptiveTreeMutation::with_probs(1.0, 0.0, 0.0, 0.0);

        for _ in 0..50 {
            let mut genome: TreeGenome<ArithmeticTerminal, ArithmeticFunction> =
                TreeGenome::generate_ramped_half_and_half(&mut rng, 2, 5);
            let original_size = genome.size();
            mutation.mutate(&mut genome, &mut rng);

            // Point mutation preserves structure
            assert_eq!(genome.size(), original_size);
        }
    }

    #[test]
    fn test_subtree_mutation_respects_max_depth() {
        // regression: EV-27 — SubtreeMutation must never grow a tree beyond its
        // `max_depth`. Pre-fix, the depth budget was off by one (generate_grow
        // can produce a subtree one level deeper than requested) and the
        // violation compounded across repeated mutations. Fuzz 500 mutations
        // with a large `max_subtree_depth` (to stress the budget) and assert the
        // depth invariant after every single mutation.
        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(2024);

        let max_depth = 5;
        // Start from a valid full tree at exactly `max_depth`
        // (generate_full(_, d, _) yields depth d + 1).
        let mut genome: TreeGenome<ArithmeticTerminal, ArithmeticFunction> =
            TreeGenome::generate_full(&mut rng, max_depth - 1, max_depth);
        assert!(genome.depth() <= max_depth);

        // Large subtree budget + low terminal probability => generate_grow tends
        // to emit functions, maximally exercising the off-by-one path.
        let mutation = SubtreeMutation::new()
            .with_max_depth(12)
            .with_terminal_probability(0.1);

        for i in 0..500 {
            mutation.mutate(&mut genome, &mut rng);
            assert!(
                genome.depth() <= max_depth,
                "iteration {i}: tree depth {} exceeded max_depth {max_depth}",
                genome.depth()
            );
        }
    }

    #[test]
    fn test_default_swap_mutations_actually_mutate() {
        // regression: EV-101 — SwapMutation / PermutationSwapMutation /
        // InsertMutation previously derived Default, giving num=0 (a silent
        // no-op). Default::default() must now behave like new() (num=1) and
        // actually change the genome.
        assert_eq!(SwapMutation::default().num_swaps, 1);
        assert_eq!(PermutationSwapMutation::default().num_swaps, 1);
        assert_eq!(InsertMutation::default().num_inserts, 1);

        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(99);

        // A default SwapMutation must eventually move genes (num_swaps=0 never would).
        let mut any_changed = false;
        for _ in 0..50 {
            let original = RealVector::new(vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]);
            let mut genome = original.clone();
            SwapMutation::default().mutate(&mut genome, &mut rng);
            if genome.genes() != original.genes() {
                any_changed = true;
                break;
            }
        }
        assert!(
            any_changed,
            "Default SwapMutation never mutated (num_swaps == 0?)"
        );

        // Default permutation swap: a valid perm with a moved element.
        let mut perm_changed = false;
        for _ in 0..50 {
            let original = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);
            let mut genome = original.clone();
            PermutationSwapMutation::default().mutate(&mut genome, &mut rng);
            if genome.as_slice() != original.as_slice() {
                perm_changed = true;
                break;
            }
        }
        assert!(
            perm_changed,
            "Default PermutationSwapMutation never mutated (num_swaps == 0?)"
        );

        // Default insert mutation.
        let mut insert_changed = false;
        for _ in 0..50 {
            let original = Permutation::new(vec![0, 1, 2, 3, 4, 5, 6, 7]);
            let mut genome = original.clone();
            InsertMutation::default().mutate(&mut genome, &mut rng);
            if genome.as_slice() != original.as_slice() {
                insert_changed = true;
                break;
            }
        }
        assert!(
            insert_changed,
            "Default InsertMutation never mutated (num_inserts == 0?)"
        );
    }

    #[test]
    fn test_unbounded_polynomial_mutation_stays_local() {
        // regression: EV-102 — the unbounded PolynomialMutation path used to
        // fabricate +/-1e10 bounds, turning a mutation into a near-random reset
        // spanning ~1e10 in magnitude. It must instead apply a local Gaussian
        // perturbation and keep genes near their original values.
        use rand::SeedableRng;
        let mut rng = rand::rngs::StdRng::seed_from_u64(1);

        for _ in 0..200 {
            let mut genome = RealVector::new(vec![0.0, 1.0, -1.0, 2.5, -3.0]);
            let mutation = PolynomialMutation::new(20.0).with_probability(1.0);
            mutation.mutate(&mut genome, &mut rng); // unbounded path

            for &g in genome.genes() {
                assert!(
                    g.abs() < 100.0,
                    "unbounded polynomial mutation produced a destructive value: {g}"
                );
            }
        }

        // A fixed sigma is honored.
        let mut genome = RealVector::new(vec![0.0; 1000]);
        let mutation = PolynomialMutation::new(20.0)
            .with_probability(1.0)
            .with_unbounded_sigma(0.05);
        mutation.mutate(&mut genome, &mut rng);
        let variance: f64 =
            genome.genes().iter().map(|g| g * g).sum::<f64>() / genome.dimension() as f64;
        // Sample std should be near 0.05 (well under any 1e10 fabrication).
        assert!(
            variance.sqrt() < 0.2,
            "fixed sigma not honored: std {}",
            variance.sqrt()
        );
    }

    #[test]
    fn test_mutation_probability_reports_effective_rate() {
        // regression: EV-103 — mutation_probability() used to report 1.0 while
        // the operators actually applied the per-gene default 1/n. It must now
        // report `None` for the length-dependent default and `Some(p)` for a
        // configured rate.
        assert_eq!(
            MutationOperator::<RealVector>::mutation_probability(&PolynomialMutation::new(20.0)),
            None
        );
        assert_eq!(
            MutationOperator::<RealVector>::mutation_probability(
                &PolynomialMutation::new(20.0).with_probability(0.25)
            ),
            Some(0.25)
        );
        assert_eq!(
            MutationOperator::<RealVector>::mutation_probability(&GaussianMutation::new(0.1)),
            None
        );
        assert_eq!(
            MutationOperator::<RealVector>::mutation_probability(&UniformMutation::new()),
            None
        );
        assert_eq!(
            MutationOperator::<BitString>::mutation_probability(&BitFlipMutation::new()),
            None
        );
        assert_eq!(
            MutationOperator::<BitString>::mutation_probability(
                &BitFlipMutation::new().with_probability(0.5)
            ),
            Some(0.5)
        );
    }
}