swarm_engine_core/exploration/

llm_provider.rs

//! LLM-based OperatorProvider - LLM による探索戦略の動的決定
//!
//! AdaptiveOperatorProvider をベースに、LLM による戦略レビューを組み合わせた
//! AdaptiveLlmOperatorProvider を提供する。
//!
//! # 設計
//!
//! ```text
//! AdaptiveLlmOperatorProvider
//! ├── AdaptiveOperatorProvider (ベースのルールロジック)
//! └── StrategyAdvisor (LLM 戦略レビュー)
//!
//! 1. 通常: AdaptiveOperatorProvider のルールベース判断
//! 2. レビュー条件を満たした時: LLM に問い合わせ（同期ブロック）
//! 3. LLM が変更を推奨 + 信頼度が閾値以上: 戦略を切り替え
//! 4. LLM エラー時: AdaptiveOperatorProvider にフォールバック
//! ```
//!
//! # 使用例
//!
//! ```ignore
//! use swarm_engine_core::exploration::{AdaptiveLlmOperatorProvider, ReviewPolicy};
//! use swarm_engine_llm::LlmStrategyAdvisor;
//!
//! let advisor = LlmStrategyAdvisor::new(decider, runtime);
//! let provider = AdaptiveLlmOperatorProvider::new(Box::new(advisor))
//!     .with_policy(ReviewPolicy::default());
//!
//! let orchestrator = OrchestratorBuilder::new()
//!     .operator_provider(provider)
//!     .build(runtime);
//! ```

use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::RwLock;
use std::time::Instant;

use super::map::MapNodeState;
use super::mutation::ActionNodeData;
use super::node_rules::Rules;
use super::operator::{ConfigurableOperator, Operator, RulesBasedMutation};
use super::provider::{AdaptiveOperatorProvider, OperatorProvider, ProviderContext};
use super::selection::{AnySelection, SelectionKind};
use crate::events::{LearningEvent, LearningEventChannel};
use crate::online_stats::SwarmStats;

// ============================================================================
// StrategyContext - 戦略判断に必要なコンテキスト
// ============================================================================

/// 戦略判断に必要なコンテキスト
#[derive(Debug, Clone)]
pub struct StrategyContext {
    /// フロンティア数
    pub frontier_count: usize,
    /// 総訪問数
    pub total_visits: u32,
    /// 失敗率 (0.0-1.0)
    pub failure_rate: f64,
    /// 成功率 (0.0-1.0)
    pub success_rate: f64,
    /// 現在の戦略
    pub current_strategy: SelectionKind,
    /// 探索の深さ（平均）- オプション
    pub avg_depth: Option<f32>,
}

impl StrategyContext {
    /// 新しい StrategyContext を作成
    pub fn new(
        frontier_count: usize,
        total_visits: u32,
        failure_rate: f64,
        current_strategy: SelectionKind,
    ) -> Self {
        Self {
            frontier_count,
            total_visits,
            failure_rate,
            success_rate: 1.0 - failure_rate,
            current_strategy,
            avg_depth: None,
        }
    }

    /// SwarmStats と ProviderContext から構築
    pub fn from_provider_context(
        ctx: &ProviderContext<'_, ActionNodeData, String, MapNodeState>,
        current: SelectionKind,
    ) -> Self {
        Self {
            frontier_count: ctx.frontier_count(),
            total_visits: ctx.total_visits(),
            failure_rate: ctx.stats.failure_rate(),
            success_rate: ctx.stats.success_rate(),
            current_strategy: current,
            avg_depth: None,
        }
    }

    /// SwarmStats から構築
    pub fn from_stats(stats: &SwarmStats, frontier_count: usize, current: SelectionKind) -> Self {
        Self {
            frontier_count,
            total_visits: stats.total_visits(),
            failure_rate: stats.failure_rate(),
            success_rate: stats.success_rate(),
            current_strategy: current,
            avg_depth: None,
        }
    }

    /// 平均深度を設定
    pub fn with_avg_depth(mut self, depth: f32) -> Self {
        self.avg_depth = Some(depth);
        self
    }
}

// ============================================================================
// StrategyAdvice - 戦略アドバイスの結果
// ============================================================================

/// 戦略アドバイスの結果
#[derive(Debug, Clone)]
pub struct StrategyAdvice {
    /// 推奨する戦略
    pub recommended: SelectionKind,
    /// 変更すべきか（現状維持なら false）
    pub should_change: bool,
    /// 理由（デバッグ用）
    pub reason: String,
    /// 信頼度 (0.0-1.0)
    pub confidence: f64,
}

impl StrategyAdvice {
    /// 変更なしのアドバイスを作成
    pub fn no_change(current: SelectionKind, reason: impl Into<String>) -> Self {
        Self {
            recommended: current,
            should_change: false,
            reason: reason.into(),
            confidence: 1.0,
        }
    }

    /// 変更ありのアドバイスを作成
    pub fn change_to(new: SelectionKind, reason: impl Into<String>, confidence: f64) -> Self {
        Self {
            recommended: new,
            should_change: true,
            reason: reason.into(),
            confidence,
        }
    }
}

// ============================================================================
// StrategyAdviceError
// ============================================================================

/// 戦略アドバイスエラー
#[derive(Debug, Clone, thiserror::Error)]
pub enum StrategyAdviceError {
    /// LLM 呼び出しエラー
    #[error("LLM call failed: {0}")]
    LlmError(String),
    /// レスポンスパースエラー
    #[error("Failed to parse response: {0}")]
    ParseError(String),
    /// アドバイザー利用不可
    #[error("Advisor not available")]
    Unavailable,
}

// ============================================================================
// StrategyAdvisor trait
// ============================================================================

/// 戦略アドバイザー trait
///
/// 探索状態を受け取り、最適な戦略を推奨する。
/// 同期インターフェース（SLM ローカル前提、~100ms）。
pub trait StrategyAdvisor: Send + Sync {
    /// 戦略をレビューしてアドバイスを返す
    fn advise(&self, context: &StrategyContext) -> Result<StrategyAdvice, StrategyAdviceError>;

    /// アドバイザー名
    fn name(&self) -> &str;
}

// ============================================================================
// ReviewPolicy - レビュー条件の設定
// ============================================================================

/// レビュー条件の設定
#[derive(Debug, Clone)]
pub struct ReviewPolicy {
    /// レビュー間隔（visits 単位）
    pub interval: u32,
    /// 最小レビュー間隔（連続レビュー防止）
    pub min_interval: u32,
    /// 状態変化でトリガーする閾値（failure_rate の変化量）
    pub state_change_threshold: f64,
}

impl Default for ReviewPolicy {
    fn default() -> Self {
        Self {
            interval: 20,                 // 20 visits 毎
            min_interval: 5,              // 最低5回は間隔を空ける
            state_change_threshold: 0.15, // 15%以上の変化でトリガー
        }
    }
}

impl ReviewPolicy {
    /// 新しい ReviewPolicy を作成
    pub fn new(interval: u32, min_interval: u32, state_change_threshold: f64) -> Self {
        Self {
            interval,
            min_interval,
            state_change_threshold: state_change_threshold.clamp(0.0, 1.0),
        }
    }

    /// 頻繁にレビューする設定
    pub fn frequent() -> Self {
        Self {
            interval: 10,
            min_interval: 3,
            state_change_threshold: 0.1,
        }
    }

    /// 控えめにレビューする設定
    pub fn conservative() -> Self {
        Self {
            interval: 50,
            min_interval: 20,
            state_change_threshold: 0.25,
        }
    }
}

// ============================================================================
// AdaptiveLlmOperatorProvider - LLM と Adaptive のハイブリッド Provider
// ============================================================================

/// Adaptive LLM Operator Provider
///
/// AdaptiveOperatorProvider をベースに、LLM による戦略レビューを組み合わせる。
///
/// # 動作
///
/// 1. 通常: AdaptiveOperatorProvider のルールベース判断
/// 2. レビュー条件を満たした時: LLM に問い合わせ（同期ブロック、~100ms）
/// 3. LLM が変更を推奨 + 信頼度が閾値以上: 戦略を切り替え
/// 4. LLM エラー時: AdaptiveOperatorProvider にフォールバック
pub struct AdaptiveLlmOperatorProvider {
    /// ベースの Adaptive ロジック
    adaptive: AdaptiveOperatorProvider,
    /// 戦略アドバイザー
    advisor: Box<dyn StrategyAdvisor>,
    /// レビューポリシー
    policy: ReviewPolicy,
    /// UCB1 の探索係数
    ucb1_c: f64,
    /// 最後のレビュー時点の visits
    last_review_visits: AtomicU32,
    /// 最後のレビュー時点の failure_rate（f64 * 1000 で保存）
    last_failure_rate: AtomicU32,
    /// LLM によるオーバーライド（None = Adaptive に従う）
    llm_override: RwLock<Option<SelectionKind>>,
}

impl AdaptiveLlmOperatorProvider {
    /// 新しい AdaptiveLlmOperatorProvider を作成
    pub fn new(advisor: Box<dyn StrategyAdvisor>) -> Self {
        Self {
            adaptive: AdaptiveOperatorProvider::default(),
            advisor,
            policy: ReviewPolicy::default(),
            ucb1_c: std::f64::consts::SQRT_2,
            last_review_visits: AtomicU32::new(0),
            last_failure_rate: AtomicU32::new(0),
            llm_override: RwLock::new(None),
        }
    }

    /// レビューポリシーを設定
    pub fn with_policy(mut self, policy: ReviewPolicy) -> Self {
        self.policy = policy;
        self
    }

    /// ベースの AdaptiveOperatorProvider を設定
    pub fn with_adaptive(mut self, adaptive: AdaptiveOperatorProvider) -> Self {
        self.adaptive = adaptive;
        self
    }

    /// UCB1 の探索係数を設定
    pub fn with_ucb1_c(mut self, c: f64) -> Self {
        self.ucb1_c = c;
        self
    }

    /// 現在の LLM override を取得
    pub fn llm_override(&self) -> Option<SelectionKind> {
        *self.llm_override.read().unwrap()
    }

    /// レビューが必要か判定
    fn should_review(&self, stats: &SwarmStats) -> bool {
        let current_visits = stats.total_visits();
        let last_visits = self.last_review_visits.load(Ordering::Relaxed);

        // 最小間隔チェック
        if current_visits < last_visits + self.policy.min_interval {
            return false;
        }

        // 定期レビュー
        if current_visits >= last_visits + self.policy.interval {
            return true;
        }

        // 状態変化チェック（failure_rate の急変）
        let current_rate = (stats.failure_rate() * 1000.0) as u32;
        let last_rate = self.last_failure_rate.load(Ordering::Relaxed);
        let rate_diff = (current_rate as i32 - last_rate as i32).unsigned_abs() as f64 / 1000.0;

        rate_diff >= self.policy.state_change_threshold
    }

    /// LLM レビューを実行
    fn do_review(
        &self,
        stats: &SwarmStats,
        frontier_count: usize,
        current: SelectionKind,
    ) -> Option<SelectionKind> {
        let context = StrategyContext::from_stats(stats, frontier_count, current);

        // タイミング計測開始
        let start_time = Instant::now();

        // 同期ブロッキング呼び出し（~100ms 想定）
        let result = self.advisor.advise(&context);

        // タイミング計測終了
        let elapsed = start_time.elapsed();

        match result {
            Ok(advice) => {
                let latency_ms = elapsed.as_millis() as u64;

                // 学習記録: LearningEventChannel に emit
                let tick = LearningEventChannel::global().current_tick();
                LearningEventChannel::global().emit(
                    LearningEvent::strategy_advice(tick, self.advisor.name())
                        .current_strategy(current.to_string())
                        .recommended(advice.recommended.to_string())
                        .should_change(advice.should_change)
                        .confidence(advice.confidence)
                        .reason(&advice.reason)
                        .frontier_count(frontier_count)
                        .total_visits(stats.total_visits())
                        .failure_rate(stats.failure_rate())
                        .latency_ms(latency_ms)
                        .success()
                        .build(),
                );

                // 構造化ログも出力（デバッグ/監視用）
                tracing::debug!(
                    target: "swarm_engine::learning",
                    advisor = %self.advisor.name(),
                    current_strategy = %current,
                    recommended = %advice.recommended,
                    should_change = advice.should_change,
                    confidence = advice.confidence,
                    reason = %advice.reason,
                    latency_ms = latency_ms,
                    "Strategy advice completed"
                );

                // レビュー状態を更新
                self.last_review_visits
                    .store(stats.total_visits(), Ordering::Relaxed);
                self.last_failure_rate
                    .store((stats.failure_rate() * 1000.0) as u32, Ordering::Relaxed);

                if advice.should_change {
                    Some(advice.recommended)
                } else {
                    None
                }
            }
            Err(e) => {
                let latency_ms = elapsed.as_millis() as u64;

                // 学習記録: LearningEventChannel に emit（エラー時）
                let tick = LearningEventChannel::global().current_tick();
                LearningEventChannel::global().emit(
                    LearningEvent::strategy_advice(tick, self.advisor.name())
                        .current_strategy(current.to_string())
                        .recommended(current.to_string()) // エラー時は変更なし
                        .frontier_count(frontier_count)
                        .total_visits(stats.total_visits())
                        .failure_rate(stats.failure_rate())
                        .latency_ms(latency_ms)
                        .failure(e.to_string())
                        .build(),
                );

                // 構造化ログ（エラー）
                tracing::warn!(
                    advisor = %self.advisor.name(),
                    error = %e,
                    latency_ms = latency_ms,
                    "Strategy advisor failed, falling back to Adaptive"
                );
                None
            }
        }
    }

    /// 現在有効な SelectionKind を取得
    fn effective_selection(&self, stats: &SwarmStats) -> SelectionKind {
        // LLM override があればそれを使用
        if let Some(kind) = *self.llm_override.read().unwrap() {
            return kind;
        }
        // なければ Adaptive のルール
        self.adaptive.current_selection(stats)
    }
}

impl<R> OperatorProvider<R> for AdaptiveLlmOperatorProvider
where
    R: Rules + 'static,
{
    fn provide(
        &self,
        rules: R,
        context: Option<&ProviderContext<'_, ActionNodeData, String, MapNodeState>>,
    ) -> ConfigurableOperator<R> {
        let selection_kind = match context {
            Some(ctx) => {
                let current = self.effective_selection(ctx.stats);

                // レビュー条件を満たしていれば LLM に問い合わせ
                if self.should_review(ctx.stats) {
                    if let Some(new_kind) = self.do_review(ctx.stats, ctx.frontier_count(), current)
                    {
                        *self.llm_override.write().unwrap() = Some(new_kind);
                        new_kind
                    } else {
                        current
                    }
                } else {
                    current
                }
            }
            None => SelectionKind::Ucb1, // 初回は UCB1
        };

        let selection = AnySelection::from_kind(selection_kind, self.ucb1_c);
        Operator::new(RulesBasedMutation::new(), selection, rules)
    }

    fn reevaluate(
        &self,
        operator: &mut ConfigurableOperator<R>,
        context: &ProviderContext<'_, ActionNodeData, String, MapNodeState>,
    ) {
        let current = operator.selection().kind();

        // レビュー条件を満たしていれば LLM に問い合わせ
        if self.should_review(context.stats) {
            if let Some(new_kind) = self.do_review(context.stats, context.frontier_count(), current)
            {
                if new_kind != current {
                    tracing::info!(
                        from = %current,
                        to = %new_kind,
                        "Strategy changed by LLM advisor"
                    );
                    operator.set_selection(AnySelection::from_kind(new_kind, self.ucb1_c));
                    *self.llm_override.write().unwrap() = Some(new_kind);
                }
                return;
            }
        }

        // LLM が変更しない場合は Adaptive のルールを適用
        // ただし llm_override がある場合はそれを維持
        if self.llm_override.read().unwrap().is_none() {
            self.adaptive.reevaluate(operator, context);
        }
    }

    fn name(&self) -> &str {
        "HybridLlm"
    }
}

// ============================================================================
// Tests
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;
    use crate::events::{ActionEventBuilder, ActionEventResult};
    use crate::exploration::{GraphMap, NodeRules};
    use crate::types::WorkerId;

    fn record_success(stats: &mut SwarmStats, action: &str) {
        let event = ActionEventBuilder::new(0, WorkerId(0), action)
            .result(ActionEventResult::success())
            .build();
        stats.record(&event);
    }

    fn record_failure(stats: &mut SwarmStats, action: &str) {
        let event = ActionEventBuilder::new(0, WorkerId(0), action)
            .result(ActionEventResult::failure("error"))
            .build();
        stats.record(&event);
    }

    // ========================================================================
    // Mock StrategyAdvisor for testing
    // ========================================================================

    struct MockAdvisor {
        advice: StrategyAdvice,
        call_count: std::sync::atomic::AtomicUsize,
    }

    impl MockAdvisor {
        fn new(advice: StrategyAdvice) -> Self {
            Self {
                advice,
                call_count: std::sync::atomic::AtomicUsize::new(0),
            }
        }

        fn call_count(&self) -> usize {
            self.call_count.load(Ordering::Relaxed)
        }
    }

    impl StrategyAdvisor for MockAdvisor {
        fn advise(
            &self,
            _context: &StrategyContext,
        ) -> Result<StrategyAdvice, StrategyAdviceError> {
            self.call_count
                .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
            Ok(self.advice.clone())
        }

        fn name(&self) -> &str {
            "MockAdvisor"
        }
    }

    struct FailingAdvisor;

    impl StrategyAdvisor for FailingAdvisor {
        fn advise(
            &self,
            _context: &StrategyContext,
        ) -> Result<StrategyAdvice, StrategyAdviceError> {
            Err(StrategyAdviceError::LlmError("Mock error".into()))
        }

        fn name(&self) -> &str {
            "FailingAdvisor"
        }
    }

    // ========================================================================
    // StrategyContext Tests
    // ========================================================================

    #[test]
    fn test_strategy_context_new() {
        let ctx = StrategyContext::new(15, 47, 0.23, SelectionKind::Ucb1);
        assert_eq!(ctx.frontier_count, 15);
        assert_eq!(ctx.total_visits, 47);
        assert!((ctx.failure_rate - 0.23).abs() < 0.001);
        assert!((ctx.success_rate - 0.77).abs() < 0.001);
        assert_eq!(ctx.current_strategy, SelectionKind::Ucb1);
    }

    #[test]
    fn test_strategy_context_from_stats() {
        let mut stats = SwarmStats::new();
        for _ in 0..7 {
            record_success(&mut stats, "action");
        }
        for _ in 0..3 {
            record_failure(&mut stats, "action");
        }

        let ctx = StrategyContext::from_stats(&stats, 10, SelectionKind::Greedy);
        assert_eq!(ctx.frontier_count, 10);
        assert_eq!(ctx.total_visits, 10);
        assert!((ctx.failure_rate - 0.3).abs() < 0.01);
    }

    // ========================================================================
    // ReviewPolicy Tests
    // ========================================================================

    #[test]
    fn test_review_policy_default() {
        let policy = ReviewPolicy::default();
        assert_eq!(policy.interval, 20);
        assert_eq!(policy.min_interval, 5);
        assert!((policy.state_change_threshold - 0.15).abs() < 0.001);
    }

    #[test]
    fn test_review_policy_frequent() {
        let policy = ReviewPolicy::frequent();
        assert_eq!(policy.interval, 10);
        assert_eq!(policy.min_interval, 3);
    }

    #[test]
    fn test_review_policy_conservative() {
        let policy = ReviewPolicy::conservative();
        assert_eq!(policy.interval, 50);
        assert_eq!(policy.min_interval, 20);
    }

    // ========================================================================
    // AdaptiveLlmOperatorProvider Tests
    // ========================================================================

    #[test]
    fn test_hybrid_provider_initial_ucb1() {
        let advice = StrategyAdvice::no_change(SelectionKind::Ucb1, "test");
        let advisor = MockAdvisor::new(advice);
        let provider = AdaptiveLlmOperatorProvider::new(Box::new(advisor));
        let rules = NodeRules::for_testing();

        // Context なしで構築 → UCB1
        let operator = provider.provide(rules, None);
        assert_eq!(operator.name(), "RulesBased+UCB1");
    }

    #[test]
    fn test_hybrid_provider_review_at_interval() {
        let advice = StrategyAdvice::change_to(SelectionKind::Greedy, "test", 0.9);
        let advisor = MockAdvisor::new(advice);
        let provider =
            AdaptiveLlmOperatorProvider::new(Box::new(advisor)).with_policy(ReviewPolicy {
                interval: 10,
                min_interval: 5,
                state_change_threshold: 0.5,
            });
        let rules = NodeRules::for_testing();

        // stats を作成（20 visits）
        let mut stats = SwarmStats::new();
        for _ in 0..20 {
            record_success(&mut stats, "action");
        }

        let map: GraphMap<ActionNodeData, String, MapNodeState> = GraphMap::new();
        let ctx = ProviderContext::new(&map, &stats);

        // レビュー条件（interval=10）を満たしているので LLM に問い合わせ
        let operator = provider.provide(rules, Some(&ctx));
        assert_eq!(operator.name(), "RulesBased+Greedy");

        // advisor が呼ばれた
        let advisor_ref = provider.advisor.as_ref();
        let mock = unsafe { &*(advisor_ref as *const dyn StrategyAdvisor as *const MockAdvisor) };
        assert_eq!(mock.call_count(), 1);
    }

    #[test]
    fn test_hybrid_provider_no_review_before_min_interval() {
        let advice = StrategyAdvice::change_to(SelectionKind::Greedy, "test", 0.9);
        let advisor = MockAdvisor::new(advice);
        let provider =
            AdaptiveLlmOperatorProvider::new(Box::new(advisor)).with_policy(ReviewPolicy {
                interval: 10,
                min_interval: 5,
                state_change_threshold: 0.5,
            });
        let rules = NodeRules::for_testing();

        // stats を作成（3 visits = min_interval 未満）
        let mut stats = SwarmStats::new();
        for _ in 0..3 {
            record_success(&mut stats, "action");
        }

        let map: GraphMap<ActionNodeData, String, MapNodeState> = GraphMap::new();
        let ctx = ProviderContext::new(&map, &stats);

        // min_interval 未満なのでレビューしない → Adaptive のルール（UCB1）
        let operator = provider.provide(rules, Some(&ctx));
        assert_eq!(operator.name(), "RulesBased+UCB1");
    }

    #[test]
    fn test_hybrid_provider_fallback_on_error() {
        let provider =
            AdaptiveLlmOperatorProvider::new(Box::new(FailingAdvisor)).with_policy(ReviewPolicy {
                interval: 1,
                min_interval: 1,
                state_change_threshold: 0.0,
            });
        let rules = NodeRules::for_testing();

        let mut stats = SwarmStats::new();
        for _ in 0..10 {
            record_success(&mut stats, "action");
        }

        let map: GraphMap<ActionNodeData, String, MapNodeState> = GraphMap::new();
        let ctx = ProviderContext::new(&map, &stats);

        // advisor がエラーを返すが、フォールバックで動作
        let operator = provider.provide(rules, Some(&ctx));
        // Adaptive のルール（成熟 + 低エラー → Greedy）が適用される
        assert!(operator.name().contains("RulesBased"));
    }

    #[test]
    fn test_hybrid_provider_reevaluate() {
        let advice = StrategyAdvice::change_to(SelectionKind::Thompson, "high variance", 0.85);
        let advisor = MockAdvisor::new(advice);
        let provider =
            AdaptiveLlmOperatorProvider::new(Box::new(advisor)).with_policy(ReviewPolicy {
                interval: 5,
                min_interval: 1,
                state_change_threshold: 0.5,
            });
        let rules = NodeRules::for_testing();

        // 初期状態で構築
        let mut operator = provider.provide(rules, None);
        assert_eq!(operator.selection().kind(), SelectionKind::Ucb1);

        // stats を更新
        let mut stats = SwarmStats::new();
        for _ in 0..10 {
            record_success(&mut stats, "action");
        }
        let map: GraphMap<ActionNodeData, String, MapNodeState> = GraphMap::new();
        let ctx = ProviderContext::new(&map, &stats);

        // reevaluate で Thompson に切り替わる
        provider.reevaluate(&mut operator, &ctx);
        assert_eq!(operator.selection().kind(), SelectionKind::Thompson);
    }

    #[test]
    fn test_hybrid_provider_state_change_trigger() {
        let advice = StrategyAdvice::change_to(SelectionKind::Thompson, "high variance", 0.8);
        let advisor = MockAdvisor::new(advice);
        let provider =
            AdaptiveLlmOperatorProvider::new(Box::new(advisor)).with_policy(ReviewPolicy {
                interval: 100, // 高い interval
                min_interval: 1,
                state_change_threshold: 0.1, // 10% の変化でトリガー
            });
        let rules = NodeRules::for_testing();

        // 最初のレビュー（初期化）
        let mut stats = SwarmStats::new();
        for _ in 0..5 {
            record_success(&mut stats, "action");
        }
        let map: GraphMap<ActionNodeData, String, MapNodeState> = GraphMap::new();
        let ctx = ProviderContext::new(&map, &stats);
        let _ = provider.provide(rules.clone(), Some(&ctx));

        // failure_rate が急変（0% → 50%）
        for _ in 0..5 {
            record_failure(&mut stats, "action");
        }
        let ctx2 = ProviderContext::new(&map, &stats);

        // interval 未到達だが、state_change_threshold を超えているのでレビュー
        let operator = provider.provide(rules, Some(&ctx2));
        assert_eq!(operator.selection().kind(), SelectionKind::Thompson);
    }
}