rsigma_eval/

correlation_engine.rs

//! Stateful correlation engine with time-windowed aggregation.
//!
//! `CorrelationEngine` wraps the stateless `Engine` and adds support for
//! Sigma correlation rules: `event_count`, `value_count`, `temporal`,
//! `temporal_ordered`, `value_sum`, `value_avg`, `value_percentile`,
//! and `value_median`.
//!
//! # Architecture
//!
//! 1. Events are first evaluated against detection rules (stateless)
//! 2. Detection matches update correlation window state (stateful)
//! 3. When a correlation condition is met, a `CorrelationResult` is emitted
//! 4. Correlation results can chain into higher-level correlations

use std::collections::HashMap;

use chrono::{DateTime, TimeZone, Utc};
use serde::Serialize;

use rsigma_parser::{CorrelationRule, CorrelationType, Level, SigmaCollection, SigmaRule};

use crate::correlation::{
    CompiledCorrelation, EventBuffer, EventRef, EventRefBuffer, GroupKey, WindowState,
    compile_correlation,
};
use crate::engine::Engine;
use crate::error::Result;
use crate::event::Event;
use crate::pipeline::{Pipeline, apply_pipelines};
use crate::result::MatchResult;

// =============================================================================
// Configuration
// =============================================================================

/// What to do with window state after a correlation fires.
///
/// This is an engine-level default that can be overridden per-correlation
/// via the `rsigma.action` custom attribute set in processing pipelines.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize)]
#[serde(rename_all = "snake_case")]
pub enum CorrelationAction {
    /// Keep window state as-is after firing (current / default behavior).
    /// Subsequent events that still satisfy the condition will re-fire.
    #[default]
    Alert,
    /// Clear the window state for the firing group key after emitting the alert.
    /// The threshold must be met again from scratch before the next alert.
    Reset,
}

impl std::str::FromStr for CorrelationAction {
    type Err = String;
    fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
        match s {
            "alert" => Ok(CorrelationAction::Alert),
            "reset" => Ok(CorrelationAction::Reset),
            _ => Err(format!(
                "Unknown correlation action: {s} (expected 'alert' or 'reset')"
            )),
        }
    }
}

/// How to include events in correlation results.
///
/// Can be overridden per-correlation via the `rsigma.correlation_event_mode`
/// custom attribute.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize)]
#[serde(rename_all = "snake_case")]
pub enum CorrelationEventMode {
    /// Don't include events (default). Zero memory overhead.
    #[default]
    None,
    /// Include full event bodies, individually compressed with deflate.
    /// Typical cost: 100–1000 bytes per event.
    Full,
    /// Include only event references (timestamp + optional ID).
    /// Minimal memory: ~40 bytes per event.
    Refs,
}

impl std::str::FromStr for CorrelationEventMode {
    type Err = String;
    fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
        match s.to_lowercase().as_str() {
            "none" | "off" | "false" => Ok(CorrelationEventMode::None),
            "full" | "true" => Ok(CorrelationEventMode::Full),
            "refs" | "references" => Ok(CorrelationEventMode::Refs),
            _ => Err(format!(
                "Unknown correlation event mode: {s} (expected 'none', 'full', or 'refs')"
            )),
        }
    }
}

impl std::fmt::Display for CorrelationEventMode {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            CorrelationEventMode::None => write!(f, "none"),
            CorrelationEventMode::Full => write!(f, "full"),
            CorrelationEventMode::Refs => write!(f, "refs"),
        }
    }
}

/// Behavior when no timestamp field is found or parseable in an event.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum TimestampFallback {
    /// Use wall-clock time (`Utc::now()`). Good for real-time streaming.
    #[default]
    WallClock,
    /// Skip the event from correlation processing. Detections still fire,
    /// but the event does not update any correlation state. Recommended for
    /// batch/replay of historical logs where wall-clock time is meaningless.
    Skip,
}

/// Configuration for the correlation engine.
///
/// Provides engine-level defaults that mirror pySigma backend optional arguments.
/// Per-correlation overrides can be set via `SetCustomAttribute` pipeline
/// transformations using the `rsigma.*` attribute namespace.
#[derive(Debug, Clone)]
pub struct CorrelationConfig {
    /// Field names to try for timestamp extraction, in order of priority.
    ///
    /// The engine will try each field until one yields a parseable timestamp.
    /// If none succeed, the `timestamp_fallback` policy applies.
    pub timestamp_fields: Vec<String>,

    /// What to do when no timestamp can be extracted from an event.
    ///
    /// Default: `WallClock` (use `Utc::now()`).
    pub timestamp_fallback: TimestampFallback,

    /// Maximum number of state entries (across all correlations and groups)
    /// before aggressive eviction is triggered. Prevents unbounded memory growth.
    ///
    /// Default: 100_000.
    pub max_state_entries: usize,

    /// Default suppression window in seconds.
    ///
    /// After a correlation fires for a `(correlation, group_key)`, suppress
    /// re-alerts for this duration. `None` means no suppression (every
    /// condition-satisfying event produces an alert).
    ///
    /// Can be overridden per-correlation via the `rsigma.suppress` custom attribute.
    pub suppress: Option<u64>,

    /// Default action to take after a correlation fires.
    ///
    /// Can be overridden per-correlation via the `rsigma.action` custom attribute.
    pub action_on_match: CorrelationAction,

    /// Whether to emit detection-level matches for rules that are only
    /// referenced by correlations (where `generate: false`).
    ///
    /// Default: `true` (emit all detection matches).
    /// Set to `false` to suppress detection output for correlation-only rules.
    pub emit_detections: bool,

    /// How to include contributing events in correlation results.
    ///
    /// - `None` (default): no event storage, zero overhead.
    /// - `Full`: events are deflate-compressed and decompressed on output.
    /// - `Refs`: only timestamps + event IDs are stored (minimal memory).
    ///
    /// Can be overridden per-correlation via `rsigma.correlation_event_mode`.
    pub correlation_event_mode: CorrelationEventMode,

    /// Maximum number of events to store per (correlation, group_key) window
    /// when `correlation_event_mode` is not `None`.
    ///
    /// Bounds memory at: `max_correlation_events × cost_per_event × active_groups`.
    /// Default: 10.
    pub max_correlation_events: usize,
}

impl Default for CorrelationConfig {
    fn default() -> Self {
        CorrelationConfig {
            timestamp_fields: vec![
                "@timestamp".to_string(),
                "timestamp".to_string(),
                "EventTime".to_string(),
                "TimeCreated".to_string(),
                "eventTime".to_string(),
            ],
            timestamp_fallback: TimestampFallback::default(),
            max_state_entries: 100_000,
            suppress: None,
            action_on_match: CorrelationAction::default(),
            emit_detections: true,
            correlation_event_mode: CorrelationEventMode::default(),
            max_correlation_events: 10,
        }
    }
}

// =============================================================================
// Result types
// =============================================================================

/// Combined result from processing a single event.
#[derive(Debug, Clone, Serialize)]
pub struct ProcessResult {
    /// Detection rule matches (stateless, immediate).
    pub detections: Vec<MatchResult>,
    /// Correlation rule matches (stateful, accumulated).
    pub correlations: Vec<CorrelationResult>,
}

/// The result of a correlation rule firing.
#[derive(Debug, Clone, Serialize)]
pub struct CorrelationResult {
    /// Title of the correlation rule.
    pub rule_title: String,
    /// ID of the correlation rule (if present).
    pub rule_id: Option<String>,
    /// Severity level.
    pub level: Option<Level>,
    /// Tags from the correlation rule.
    pub tags: Vec<String>,
    /// Type of correlation.
    pub correlation_type: CorrelationType,
    /// Group-by field names and their values for this match.
    pub group_key: Vec<(String, String)>,
    /// The aggregated value that triggered the condition (count, sum, avg, etc.).
    pub aggregated_value: f64,
    /// The time window in seconds.
    pub timespan_secs: u64,
    /// Full event bodies, included when `correlation_event_mode` is `Full`.
    ///
    /// Contains up to `max_correlation_events` recently stored window events.
    /// Events are decompressed from deflate storage on output.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub events: Option<Vec<serde_json::Value>>,
    /// Lightweight event references, included when `correlation_event_mode` is `Refs`.
    ///
    /// Contains up to `max_correlation_events` timestamp + optional ID pairs.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub event_refs: Option<Vec<EventRef>>,
}

// =============================================================================
// Correlation Engine
// =============================================================================

/// Stateful correlation engine.
///
/// Wraps the stateless `Engine` for detection rules and adds time-windowed
/// correlation on top. Supports all 7 Sigma correlation types and chaining.
pub struct CorrelationEngine {
    /// Inner stateless detection engine.
    engine: Engine,
    /// Compiled correlation rules.
    correlations: Vec<CompiledCorrelation>,
    /// Maps rule ID/name -> indices into `correlations` that reference it.
    /// This allows quick lookup: "which correlations care about rule X?"
    rule_index: HashMap<String, Vec<usize>>,
    /// Maps detection rule index -> (rule_id, rule_name) for reverse lookup.
    /// Used to find which correlations a detection match triggers.
    rule_ids: Vec<(Option<String>, Option<String>)>,
    /// Per-(correlation_index, group_key) window state.
    state: HashMap<(usize, GroupKey), WindowState>,
    /// Last alert timestamp per (correlation_index, group_key) for suppression.
    last_alert: HashMap<(usize, GroupKey), i64>,
    /// Per-(correlation_index, group_key) compressed event buffer (`Full` mode).
    event_buffers: HashMap<(usize, GroupKey), EventBuffer>,
    /// Per-(correlation_index, group_key) event reference buffer (`Refs` mode).
    event_ref_buffers: HashMap<(usize, GroupKey), EventRefBuffer>,
    /// Set of detection rule IDs/names that are "correlation-only"
    /// (referenced by correlations where `generate == false`).
    /// Used to filter detection output when `config.emit_detections == false`.
    correlation_only_rules: std::collections::HashSet<String>,
    /// Configuration.
    config: CorrelationConfig,
    /// Processing pipelines applied to rules during add_rule.
    pipelines: Vec<Pipeline>,
}

impl CorrelationEngine {
    /// Create a new correlation engine with the given configuration.
    pub fn new(config: CorrelationConfig) -> Self {
        CorrelationEngine {
            engine: Engine::new(),
            correlations: Vec::new(),
            rule_index: HashMap::new(),
            rule_ids: Vec::new(),
            state: HashMap::new(),
            last_alert: HashMap::new(),
            event_buffers: HashMap::new(),
            event_ref_buffers: HashMap::new(),
            correlation_only_rules: std::collections::HashSet::new(),
            config,
            pipelines: Vec::new(),
        }
    }

    /// Add a pipeline to the engine.
    ///
    /// Pipelines are applied to rules during `add_rule` / `add_collection`.
    pub fn add_pipeline(&mut self, pipeline: Pipeline) {
        self.pipelines.push(pipeline);
        self.pipelines.sort_by_key(|p| p.priority);
    }

    /// Set global `include_event` on the inner detection engine.
    pub fn set_include_event(&mut self, include: bool) {
        self.engine.set_include_event(include);
    }

    /// Set the global correlation event mode.
    ///
    /// - `None`: no event storage (default)
    /// - `Full`: compressed event bodies
    /// - `Refs`: lightweight timestamp + ID references
    pub fn set_correlation_event_mode(&mut self, mode: CorrelationEventMode) {
        self.config.correlation_event_mode = mode;
    }

    /// Set the maximum number of events to store per correlation window group.
    /// Only meaningful when `correlation_event_mode` is not `None`.
    pub fn set_max_correlation_events(&mut self, max: usize) {
        self.config.max_correlation_events = max;
    }

    /// Add a single detection rule.
    ///
    /// If pipelines are set, the rule is cloned and transformed before compilation.
    /// The inner engine receives the already-transformed rule directly (not through
    /// its own pipeline, to avoid double transformation).
    pub fn add_rule(&mut self, rule: &SigmaRule) -> Result<()> {
        if self.pipelines.is_empty() {
            self.apply_custom_attributes(&rule.custom_attributes);
            self.rule_ids.push((rule.id.clone(), rule.name.clone()));
            self.engine.add_rule(rule)?;
        } else {
            let mut transformed = rule.clone();
            apply_pipelines(&self.pipelines, &mut transformed)?;
            self.apply_custom_attributes(&transformed.custom_attributes);
            self.rule_ids
                .push((transformed.id.clone(), transformed.name.clone()));
            // Use compile_rule + add_compiled_rule to bypass inner engine's pipelines
            let compiled = crate::compiler::compile_rule(&transformed)?;
            self.engine.add_compiled_rule(compiled);
        }
        Ok(())
    }

    /// Read `rsigma.*` custom attributes from a rule and apply them to the
    /// engine configuration.  This allows pipelines to influence engine
    /// behaviour via `SetCustomAttribute` transformations — the same pattern
    /// used by pySigma backends (e.g. pySigma-backend-loki).
    ///
    /// Supported attributes:
    /// - `rsigma.timestamp_field` — prepends a field name to the timestamp
    ///   extraction priority list so the correlation engine can find the
    ///   event timestamp in non-standard field names.
    /// - `rsigma.suppress` — sets the default suppression window (e.g. `5m`).
    ///   Only applied when the CLI did not already set `--suppress`.
    /// - `rsigma.action` — sets the default post-fire action (`alert`/`reset`).
    ///   Only applied when the CLI did not already set `--action`.
    fn apply_custom_attributes(&mut self, attrs: &std::collections::HashMap<String, String>) {
        // rsigma.timestamp_field — prepend to priority list, skip duplicates
        if let Some(field) = attrs.get("rsigma.timestamp_field")
            && !self.config.timestamp_fields.contains(field)
        {
            self.config.timestamp_fields.insert(0, field.clone());
        }

        // rsigma.suppress — only when CLI didn't already set one
        if let Some(val) = attrs.get("rsigma.suppress")
            && self.config.suppress.is_none()
            && let Ok(ts) = rsigma_parser::Timespan::parse(val)
        {
            self.config.suppress = Some(ts.seconds);
        }

        // rsigma.action — only when CLI left it at the default (Alert)
        if let Some(val) = attrs.get("rsigma.action")
            && self.config.action_on_match == CorrelationAction::Alert
            && let Ok(a) = val.parse::<CorrelationAction>()
        {
            self.config.action_on_match = a;
        }
    }

    /// Add a single correlation rule.
    pub fn add_correlation(&mut self, corr: &CorrelationRule) -> Result<()> {
        // Apply engine-level custom attributes from the correlation rule
        // (e.g. rsigma.timestamp_field).
        self.apply_custom_attributes(&corr.custom_attributes);

        let compiled = compile_correlation(corr)?;
        let idx = self.correlations.len();

        // Index by each referenced rule ID/name
        for rule_ref in &compiled.rule_refs {
            self.rule_index
                .entry(rule_ref.clone())
                .or_default()
                .push(idx);
        }

        // Track correlation-only rules (generate == false is the default)
        if !compiled.generate {
            for rule_ref in &compiled.rule_refs {
                self.correlation_only_rules.insert(rule_ref.clone());
            }
        }

        self.correlations.push(compiled);
        Ok(())
    }

    /// Add all rules and correlations from a parsed collection.
    ///
    /// Detection rules are added first (so they're available for correlation
    /// references), then correlation rules.
    pub fn add_collection(&mut self, collection: &SigmaCollection) -> Result<()> {
        for rule in &collection.rules {
            self.add_rule(rule)?;
        }
        // Apply filter rules to the inner engine's detection rules
        for filter in &collection.filters {
            self.engine.apply_filter(filter)?;
        }
        for corr in &collection.correlations {
            self.add_correlation(corr)?;
        }
        // Validate no cycles exist in the correlation reference graph
        self.detect_correlation_cycles()?;
        Ok(())
    }

    /// Detect cycles in the correlation reference graph.
    ///
    /// Builds a directed graph where each correlation (identified by its id/name)
    /// has edges to the correlations it references via `rule_refs`. Uses DFS with
    /// a "gray/black" coloring scheme to detect back-edges (cycles).
    ///
    /// Returns `Err(EvalError::CorrelationCycle)` if a cycle is found.
    fn detect_correlation_cycles(&self) -> Result<()> {
        // Build a set of all correlation identifiers (id and/or name)
        let mut corr_identifiers: HashMap<&str, usize> = HashMap::new();
        for (idx, corr) in self.correlations.iter().enumerate() {
            if let Some(ref id) = corr.id {
                corr_identifiers.insert(id.as_str(), idx);
            }
            if let Some(ref name) = corr.name {
                corr_identifiers.insert(name.as_str(), idx);
            }
        }

        // Build adjacency list: corr index → set of corr indices it references
        let mut adj: Vec<Vec<usize>> = vec![Vec::new(); self.correlations.len()];
        for (idx, corr) in self.correlations.iter().enumerate() {
            for rule_ref in &corr.rule_refs {
                if let Some(&target_idx) = corr_identifiers.get(rule_ref.as_str()) {
                    adj[idx].push(target_idx);
                }
            }
        }

        // DFS cycle detection with three states: white (unvisited), gray (in stack), black (done)
        let mut state = vec![0u8; self.correlations.len()]; // 0=white, 1=gray, 2=black
        let mut path: Vec<usize> = Vec::new();

        for start in 0..self.correlations.len() {
            if state[start] == 0
                && let Some(cycle) = Self::dfs_find_cycle(start, &adj, &mut state, &mut path)
            {
                let names: Vec<String> = cycle
                    .iter()
                    .map(|&i| {
                        self.correlations[i]
                            .id
                            .as_deref()
                            .or(self.correlations[i].name.as_deref())
                            .unwrap_or(&self.correlations[i].title)
                            .to_string()
                    })
                    .collect();
                return Err(crate::error::EvalError::CorrelationCycle(
                    names.join(" -> "),
                ));
            }
        }
        Ok(())
    }

    /// DFS helper that returns the cycle path if a back-edge is found.
    fn dfs_find_cycle(
        node: usize,
        adj: &[Vec<usize>],
        state: &mut [u8],
        path: &mut Vec<usize>,
    ) -> Option<Vec<usize>> {
        state[node] = 1; // gray
        path.push(node);

        for &next in &adj[node] {
            if state[next] == 1 {
                // Back-edge found — extract cycle from path
                if let Some(pos) = path.iter().position(|&n| n == next) {
                    let mut cycle = path[pos..].to_vec();
                    cycle.push(next); // close the cycle
                    return Some(cycle);
                }
            }
            if state[next] == 0
                && let Some(cycle) = Self::dfs_find_cycle(next, adj, state, path)
            {
                return Some(cycle);
            }
        }

        path.pop();
        state[node] = 2; // black
        None
    }

    /// Process an event, extracting the timestamp from configured event fields.
    ///
    /// When no timestamp field is found, the `timestamp_fallback` policy applies:
    /// - `WallClock`: use `Utc::now()` (good for real-time streaming)
    /// - `Skip`: return detections only, skip correlation state updates
    pub fn process_event(&mut self, event: &Event) -> ProcessResult {
        let ts = match self.extract_event_timestamp(event) {
            Some(ts) => ts,
            None => match self.config.timestamp_fallback {
                TimestampFallback::WallClock => Utc::now().timestamp(),
                TimestampFallback::Skip => {
                    // Still run detection (stateless), but skip correlation
                    let all_detections = self.engine.evaluate(event);
                    let detections = self.filter_detections(all_detections);
                    return ProcessResult {
                        detections,
                        correlations: Vec::new(),
                    };
                }
            },
        };
        self.process_event_at(event, ts)
    }

    /// Process an event with an explicit Unix epoch timestamp (seconds).
    ///
    /// The timestamp is clamped to `[0, i64::MAX / 2]` to prevent overflow
    /// when adding timespan durations internally.
    pub fn process_event_at(&mut self, event: &Event, timestamp_secs: i64) -> ProcessResult {
        let timestamp_secs = timestamp_secs.clamp(0, i64::MAX / 2);

        // Step 1: Memory management — evict before adding new state to enforce limit
        if self.state.len() >= self.config.max_state_entries {
            self.evict_all(timestamp_secs);
        }

        // Step 2: Run stateless detection
        let all_detections = self.engine.evaluate(event);

        // Step 3: Feed detection matches into correlations
        let mut correlations = Vec::new();
        self.feed_detections(event, &all_detections, timestamp_secs, &mut correlations);

        // Step 4: Chain — correlation results may trigger higher-level correlations
        self.chain_correlations(&correlations, timestamp_secs);

        // Step 5: Filter detections by generate flag
        let detections = self.filter_detections(all_detections);

        ProcessResult {
            detections,
            correlations,
        }
    }

    /// Filter detections by the `generate` flag / `emit_detections` config.
    ///
    /// If `emit_detections` is false and some rules are correlation-only,
    /// their detection output is suppressed.
    fn filter_detections(&self, all_detections: Vec<MatchResult>) -> Vec<MatchResult> {
        if !self.config.emit_detections && !self.correlation_only_rules.is_empty() {
            all_detections
                .into_iter()
                .filter(|m| {
                    let id_match = m
                        .rule_id
                        .as_ref()
                        .is_some_and(|id| self.correlation_only_rules.contains(id));
                    !id_match
                })
                .collect()
        } else {
            all_detections
        }
    }

    /// Feed detection matches into correlation window states.
    fn feed_detections(
        &mut self,
        event: &Event,
        detections: &[MatchResult],
        ts: i64,
        out: &mut Vec<CorrelationResult>,
    ) {
        // Collect all (corr_idx, rule_id, rule_name) tuples upfront to avoid
        // borrow conflicts between self.rule_ids and self.update_correlation.
        let mut work: Vec<(usize, Option<String>, Option<String>)> = Vec::new();

        for det in detections {
            // Use the MatchResult's rule_id to find the original rule's ID/name.
            // We also look up by rule_id in our rule_ids table for the name.
            let (rule_id, rule_name) = self.find_rule_identity(det);

            // Collect correlation indices that reference this rule
            let mut corr_indices = Vec::new();
            if let Some(ref id) = rule_id
                && let Some(indices) = self.rule_index.get(id)
            {
                corr_indices.extend(indices);
            }
            if let Some(ref name) = rule_name
                && let Some(indices) = self.rule_index.get(name)
            {
                corr_indices.extend(indices);
            }

            corr_indices.sort_unstable();
            corr_indices.dedup();

            for &corr_idx in &corr_indices {
                work.push((corr_idx, rule_id.clone(), rule_name.clone()));
            }
        }

        for (corr_idx, rule_id, rule_name) in work {
            self.update_correlation(corr_idx, event, ts, &rule_id, &rule_name, out);
        }
    }

    /// Find the (id, name) for a detection match by searching our rule_ids table.
    fn find_rule_identity(&self, det: &MatchResult) -> (Option<String>, Option<String>) {
        // First, try to find by matching rule_id in our table
        if let Some(ref match_id) = det.rule_id {
            for (id, name) in &self.rule_ids {
                if id.as_deref() == Some(match_id.as_str()) {
                    return (id.clone(), name.clone());
                }
            }
        }
        // Fall back to using just the MatchResult's rule_id
        (det.rule_id.clone(), None)
    }

    /// Resolve the event mode for a given correlation.
    fn resolve_event_mode(&self, corr_idx: usize) -> CorrelationEventMode {
        let corr = &self.correlations[corr_idx];
        corr.event_mode
            .unwrap_or(self.config.correlation_event_mode)
    }

    /// Resolve the max events cap for a given correlation.
    fn resolve_max_events(&self, corr_idx: usize) -> usize {
        let corr = &self.correlations[corr_idx];
        corr.max_events
            .unwrap_or(self.config.max_correlation_events)
    }

    /// Update a single correlation's state and check its condition.
    fn update_correlation(
        &mut self,
        corr_idx: usize,
        event: &Event,
        ts: i64,
        rule_id: &Option<String>,
        rule_name: &Option<String>,
        out: &mut Vec<CorrelationResult>,
    ) {
        // Borrow the correlation by reference — no cloning needed.  Rust allows
        // simultaneous &self.correlations and &mut self.state / &mut self.last_alert
        // because they are disjoint struct fields.
        let corr = &self.correlations[corr_idx];
        let corr_type = corr.correlation_type;
        let timespan = corr.timespan_secs;
        let level = corr.level;
        let suppress_secs = corr.suppress_secs.or(self.config.suppress);
        let action = corr.action.unwrap_or(self.config.action_on_match);
        let event_mode = self.resolve_event_mode(corr_idx);
        let max_events = self.resolve_max_events(corr_idx);

        // Determine the rule_ref strings for alias resolution and temporal tracking.
        let mut ref_strs: Vec<&str> = Vec::new();
        if let Some(id) = rule_id.as_deref() {
            ref_strs.push(id);
        }
        if let Some(name) = rule_name.as_deref() {
            ref_strs.push(name);
        }
        let rule_ref = rule_id.as_deref().or(rule_name.as_deref()).unwrap_or("");

        // Extract group key
        let group_key = GroupKey::extract(event, &corr.group_by, &ref_strs);

        // Get or create window state
        let state_key = (corr_idx, group_key.clone());
        let state = self
            .state
            .entry(state_key.clone())
            .or_insert_with(|| WindowState::new_for(corr_type));

        // Evict expired entries
        let cutoff = ts - timespan as i64;
        state.evict(cutoff);

        // Push the new event into the state
        match corr_type {
            CorrelationType::EventCount => {
                state.push_event_count(ts);
            }
            CorrelationType::ValueCount => {
                if let Some(ref field_name) = corr.condition.field
                    && let Some(val) = event.get_field(field_name)
                    && let Some(s) = value_to_string_for_count(val)
                {
                    state.push_value_count(ts, s);
                }
            }
            CorrelationType::Temporal | CorrelationType::TemporalOrdered => {
                state.push_temporal(ts, rule_ref);
            }
            CorrelationType::ValueSum
            | CorrelationType::ValueAvg
            | CorrelationType::ValuePercentile
            | CorrelationType::ValueMedian => {
                if let Some(ref field_name) = corr.condition.field
                    && let Some(val) = event.get_field(field_name)
                    && let Some(n) = value_to_f64(val)
                {
                    state.push_numeric(ts, n);
                }
            }
        }

        // Push event into buffer based on event mode
        match event_mode {
            CorrelationEventMode::Full => {
                let buf = self
                    .event_buffers
                    .entry(state_key.clone())
                    .or_insert_with(|| EventBuffer::new(max_events));
                buf.evict(cutoff);
                buf.push(ts, event.as_value());
            }
            CorrelationEventMode::Refs => {
                let buf = self
                    .event_ref_buffers
                    .entry(state_key.clone())
                    .or_insert_with(|| EventRefBuffer::new(max_events));
                buf.evict(cutoff);
                buf.push(ts, event.as_value());
            }
            CorrelationEventMode::None => {}
        }

        // Check condition — after this, `state` is no longer used (NLL drops the borrow).
        let fired = state.check_condition(
            &corr.condition,
            corr_type,
            &corr.rule_refs,
            corr.extended_expr.as_ref(),
        );

        if let Some(agg_value) = fired {
            let alert_key = (corr_idx, group_key.clone());

            // Suppression check: skip if we've already alerted within the suppress window
            let suppressed = if let Some(suppress) = suppress_secs {
                if let Some(&last_ts) = self.last_alert.get(&alert_key) {
                    (ts - last_ts) < suppress as i64
                } else {
                    false
                }
            } else {
                false
            };

            if !suppressed {
                // Retrieve stored events / refs based on mode
                let (events, event_refs) = match event_mode {
                    CorrelationEventMode::Full => {
                        let stored = self
                            .event_buffers
                            .get(&alert_key)
                            .map(|buf| buf.decompress_all())
                            .unwrap_or_default();
                        (Some(stored), None)
                    }
                    CorrelationEventMode::Refs => {
                        let stored = self
                            .event_ref_buffers
                            .get(&alert_key)
                            .map(|buf| buf.refs())
                            .unwrap_or_default();
                        (None, Some(stored))
                    }
                    CorrelationEventMode::None => (None, None),
                };

                // Only clone title/id/tags when we actually produce output
                let corr = &self.correlations[corr_idx];
                let result = CorrelationResult {
                    rule_title: corr.title.clone(),
                    rule_id: corr.id.clone(),
                    level,
                    tags: corr.tags.clone(),
                    correlation_type: corr_type,
                    group_key: group_key.to_pairs(&corr.group_by),
                    aggregated_value: agg_value,
                    timespan_secs: timespan,
                    events,
                    event_refs,
                };
                out.push(result);

                // Record alert time for suppression
                self.last_alert.insert(alert_key.clone(), ts);

                // Action on match
                if action == CorrelationAction::Reset {
                    if let Some(state) = self.state.get_mut(&alert_key) {
                        state.clear();
                    }
                    if let Some(buf) = self.event_buffers.get_mut(&alert_key) {
                        buf.clear();
                    }
                    if let Some(buf) = self.event_ref_buffers.get_mut(&alert_key) {
                        buf.clear();
                    }
                }
            }
        }
    }

    /// Propagate correlation results to higher-level correlations (chaining).
    ///
    /// When a correlation fires, any correlation that references it (by ID or name)
    /// is updated. Limits chain depth to 10 to prevent infinite loops.
    fn chain_correlations(&mut self, fired: &[CorrelationResult], ts: i64) {
        const MAX_CHAIN_DEPTH: usize = 10;
        let mut pending: Vec<CorrelationResult> = fired.to_vec();
        let mut depth = 0;

        while !pending.is_empty() && depth < MAX_CHAIN_DEPTH {
            depth += 1;

            // Collect work items: (corr_idx, group_key_pairs, fired_ref)
            #[allow(clippy::type_complexity)]
            let mut work: Vec<(usize, Vec<(String, String)>, String)> = Vec::new();
            for result in &pending {
                if let Some(ref id) = result.rule_id
                    && let Some(indices) = self.rule_index.get(id)
                {
                    let fired_ref = result
                        .rule_id
                        .as_deref()
                        .unwrap_or(&result.rule_title)
                        .to_string();
                    for &corr_idx in indices {
                        work.push((corr_idx, result.group_key.clone(), fired_ref.clone()));
                    }
                }
            }

            let mut next_pending = Vec::new();
            for (corr_idx, group_key_pairs, fired_ref) in work {
                let corr = &self.correlations[corr_idx];
                let corr_type = corr.correlation_type;
                let timespan = corr.timespan_secs;
                let level = corr.level;

                let group_key = GroupKey::from_pairs(&group_key_pairs, &corr.group_by);
                let state_key = (corr_idx, group_key.clone());
                let state = self
                    .state
                    .entry(state_key)
                    .or_insert_with(|| WindowState::new_for(corr_type));

                let cutoff = ts - timespan as i64;
                state.evict(cutoff);

                match corr_type {
                    CorrelationType::EventCount => {
                        state.push_event_count(ts);
                    }
                    CorrelationType::Temporal | CorrelationType::TemporalOrdered => {
                        state.push_temporal(ts, &fired_ref);
                    }
                    _ => {
                        state.push_event_count(ts);
                    }
                }

                let fired = state.check_condition(
                    &corr.condition,
                    corr_type,
                    &corr.rule_refs,
                    corr.extended_expr.as_ref(),
                );

                if let Some(agg_value) = fired {
                    let corr = &self.correlations[corr_idx];
                    next_pending.push(CorrelationResult {
                        rule_title: corr.title.clone(),
                        rule_id: corr.id.clone(),
                        level,
                        tags: corr.tags.clone(),
                        correlation_type: corr_type,
                        group_key: group_key.to_pairs(&corr.group_by),
                        aggregated_value: agg_value,
                        timespan_secs: timespan,
                        // Chained correlations don't include events (they aggregate
                        // over correlation results, not raw events)
                        events: None,
                        event_refs: None,
                    });
                }
            }

            pending = next_pending;
        }

        if !pending.is_empty() {
            log::warn!(
                "Correlation chain depth limit reached ({MAX_CHAIN_DEPTH}); \
                 {} pending result(s) were not propagated further. \
                 This may indicate a cycle in correlation references.",
                pending.len()
            );
        }
    }

    // =========================================================================
    // Timestamp extraction
    // =========================================================================

    /// Extract a Unix epoch timestamp (seconds) from an event.
    ///
    /// Tries each configured timestamp field in order. Supports:
    /// - Numeric values (epoch seconds, or epoch millis if > 1e12)
    /// - ISO 8601 strings (e.g., "2024-07-10T12:30:00Z")
    ///
    /// Returns `None` if no field yields a valid timestamp.
    fn extract_event_timestamp(&self, event: &Event) -> Option<i64> {
        for field_name in &self.config.timestamp_fields {
            if let Some(val) = event.get_field(field_name)
                && let Some(ts) = parse_timestamp_value(val)
            {
                return Some(ts);
            }
        }
        None
    }

    // =========================================================================
    // State management
    // =========================================================================

    /// Manually evict all expired state entries.
    pub fn evict_expired(&mut self, now_secs: i64) {
        self.evict_all(now_secs);
    }

    /// Evict expired entries and remove empty states.
    fn evict_all(&mut self, now_secs: i64) {
        // Phase 1: Time-based eviction — remove entries outside their correlation window
        let timespans: Vec<u64> = self.correlations.iter().map(|c| c.timespan_secs).collect();

        self.state.retain(|&(corr_idx, _), state| {
            if corr_idx < timespans.len() {
                let cutoff = now_secs - timespans[corr_idx] as i64;
                state.evict(cutoff);
            }
            !state.is_empty()
        });

        // Evict event buffers in sync with window state
        self.event_buffers.retain(|&(corr_idx, _), buf| {
            if corr_idx < timespans.len() {
                let cutoff = now_secs - timespans[corr_idx] as i64;
                buf.evict(cutoff);
            }
            !buf.is_empty()
        });
        self.event_ref_buffers.retain(|&(corr_idx, _), buf| {
            if corr_idx < timespans.len() {
                let cutoff = now_secs - timespans[corr_idx] as i64;
                buf.evict(cutoff);
            }
            !buf.is_empty()
        });

        // Phase 2: Hard cap — if still over limit after time-based eviction (e.g.
        // high-cardinality traffic with long windows), drop the stalest entries
        // until we're at 90% capacity to avoid evicting on every single event.
        if self.state.len() >= self.config.max_state_entries {
            let target = self.config.max_state_entries * 9 / 10;
            let excess = self.state.len() - target;

            // Collect keys with their latest timestamp, sort by oldest first
            let mut by_staleness: Vec<_> = self
                .state
                .iter()
                .map(|(k, v)| (k.clone(), v.latest_timestamp().unwrap_or(i64::MIN)))
                .collect();
            by_staleness.sort_unstable_by_key(|&(_, ts)| ts);

            // Drop the oldest entries (and their associated event buffers)
            for (key, _) in by_staleness.into_iter().take(excess) {
                self.state.remove(&key);
                self.last_alert.remove(&key);
                self.event_buffers.remove(&key);
                self.event_ref_buffers.remove(&key);
            }
        }

        // Phase 3: Evict stale last_alert entries — remove if the suppress window
        // has passed or if the corresponding window state no longer exists.
        self.last_alert.retain(|key, &mut alert_ts| {
            let suppress = if key.0 < self.correlations.len() {
                self.correlations[key.0]
                    .suppress_secs
                    .or(self.config.suppress)
                    .unwrap_or(0)
            } else {
                0
            };
            (now_secs - alert_ts) < suppress as i64
        });
    }

    /// Number of active state entries (for monitoring).
    pub fn state_count(&self) -> usize {
        self.state.len()
    }

    /// Number of detection rules loaded.
    pub fn detection_rule_count(&self) -> usize {
        self.engine.rule_count()
    }

    /// Number of correlation rules loaded.
    pub fn correlation_rule_count(&self) -> usize {
        self.correlations.len()
    }

    /// Number of active event buffers (for monitoring).
    pub fn event_buffer_count(&self) -> usize {
        self.event_buffers.len()
    }

    /// Total compressed bytes across all event buffers (for monitoring).
    pub fn event_buffer_bytes(&self) -> usize {
        self.event_buffers
            .values()
            .map(|b| b.compressed_bytes())
            .sum()
    }

    /// Number of active event ref buffers — `Refs` mode (for monitoring).
    pub fn event_ref_buffer_count(&self) -> usize {
        self.event_ref_buffers.len()
    }

    /// Access the inner stateless engine.
    pub fn engine(&self) -> &Engine {
        &self.engine
    }
}

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

// =============================================================================
// Timestamp parsing helpers
// =============================================================================

/// Parse a JSON value as a Unix epoch timestamp in seconds.
fn parse_timestamp_value(val: &serde_json::Value) -> Option<i64> {
    match val {
        serde_json::Value::Number(n) => {
            if let Some(i) = n.as_i64() {
                Some(normalize_epoch(i))
            } else {
                n.as_f64().map(|f| normalize_epoch(f as i64))
            }
        }
        serde_json::Value::String(s) => parse_timestamp_string(s),
        _ => None,
    }
}

/// Normalize an epoch value: if it looks like milliseconds (> year 33658),
/// convert to seconds.
fn normalize_epoch(v: i64) -> i64 {
    if v > 1_000_000_000_000 { v / 1000 } else { v }
}

/// Parse a timestamp string. Tries ISO 8601 with timezone, then without.
fn parse_timestamp_string(s: &str) -> Option<i64> {
    // Try RFC 3339 / ISO 8601 with timezone
    if let Ok(dt) = DateTime::parse_from_rfc3339(s) {
        return Some(dt.timestamp());
    }

    // Try ISO 8601 without timezone (assume UTC)
    // Common formats: "2024-07-10T12:30:00", "2024-07-10 12:30:00"
    if let Ok(naive) = chrono::NaiveDateTime::parse_from_str(s, "%Y-%m-%dT%H:%M:%S") {
        return Some(Utc.from_utc_datetime(&naive).timestamp());
    }
    if let Ok(naive) = chrono::NaiveDateTime::parse_from_str(s, "%Y-%m-%d %H:%M:%S") {
        return Some(Utc.from_utc_datetime(&naive).timestamp());
    }

    // Try with fractional seconds
    if let Ok(naive) = chrono::NaiveDateTime::parse_from_str(s, "%Y-%m-%dT%H:%M:%S%.f") {
        return Some(Utc.from_utc_datetime(&naive).timestamp());
    }
    if let Ok(naive) = chrono::NaiveDateTime::parse_from_str(s, "%Y-%m-%d %H:%M:%S%.f") {
        return Some(Utc.from_utc_datetime(&naive).timestamp());
    }

    None
}

/// Convert a JSON value to a string for value_count purposes.
fn value_to_string_for_count(v: &serde_json::Value) -> Option<String> {
    match v {
        serde_json::Value::String(s) => Some(s.clone()),
        serde_json::Value::Number(n) => Some(n.to_string()),
        serde_json::Value::Bool(b) => Some(b.to_string()),
        serde_json::Value::Null => Some("null".to_string()),
        _ => None,
    }
}

/// Convert a JSON value to f64 for numeric aggregation.
fn value_to_f64(v: &serde_json::Value) -> Option<f64> {
    match v {
        serde_json::Value::Number(n) => n.as_f64(),
        serde_json::Value::String(s) => s.parse().ok(),
        _ => None,
    }
}

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

#[cfg(test)]
mod tests {
    use super::*;
    use rsigma_parser::parse_sigma_yaml;
    use serde_json::json;

    // =========================================================================
    // Timestamp parsing
    // =========================================================================

    #[test]
    fn test_parse_timestamp_epoch_secs() {
        let val = json!(1720612200);
        assert_eq!(parse_timestamp_value(&val), Some(1720612200));
    }

    #[test]
    fn test_parse_timestamp_epoch_millis() {
        let val = json!(1720612200000i64);
        assert_eq!(parse_timestamp_value(&val), Some(1720612200));
    }

    #[test]
    fn test_parse_timestamp_rfc3339() {
        let val = json!("2024-07-10T12:30:00Z");
        let ts = parse_timestamp_value(&val).unwrap();
        assert_eq!(ts, 1720614600);
    }

    #[test]
    fn test_parse_timestamp_naive() {
        let val = json!("2024-07-10T12:30:00");
        let ts = parse_timestamp_value(&val).unwrap();
        assert_eq!(ts, 1720614600);
    }

    #[test]
    fn test_parse_timestamp_with_space() {
        let val = json!("2024-07-10 12:30:00");
        let ts = parse_timestamp_value(&val).unwrap();
        assert_eq!(ts, 1720614600);
    }

    #[test]
    fn test_parse_timestamp_fractional() {
        let val = json!("2024-07-10T12:30:00.123Z");
        let ts = parse_timestamp_value(&val).unwrap();
        assert_eq!(ts, 1720614600);
    }

    #[test]
    fn test_extract_timestamp_from_event() {
        let config = CorrelationConfig {
            timestamp_fields: vec!["@timestamp".to_string()],
            max_state_entries: 100_000,
            ..Default::default()
        };
        let engine = CorrelationEngine::new(config);

        let v = json!({"@timestamp": "2024-07-10T12:30:00Z", "data": "test"});
        let event = Event::from_value(&v);
        let ts = engine.extract_event_timestamp(&event);
        assert_eq!(ts, Some(1720614600));
    }

    #[test]
    fn test_extract_timestamp_fallback_fields() {
        let config = CorrelationConfig {
            timestamp_fields: vec![
                "@timestamp".to_string(),
                "timestamp".to_string(),
                "EventTime".to_string(),
            ],
            max_state_entries: 100_000,
            ..Default::default()
        };
        let engine = CorrelationEngine::new(config);

        // First field missing, second field present
        let v = json!({"timestamp": 1720613400, "data": "test"});
        let event = Event::from_value(&v);
        let ts = engine.extract_event_timestamp(&event);
        assert_eq!(ts, Some(1720613400));
    }

    #[test]
    fn test_extract_timestamp_returns_none_when_missing() {
        let config = CorrelationConfig {
            timestamp_fields: vec!["@timestamp".to_string()],
            ..Default::default()
        };
        let engine = CorrelationEngine::new(config);

        let v = json!({"data": "no timestamp here"});
        let event = Event::from_value(&v);
        assert_eq!(engine.extract_event_timestamp(&event), None);
    }

    #[test]
    fn test_timestamp_fallback_skip() {
        let yaml = r#"
title: test rule
id: ts-skip-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: test correlation
correlation:
    type: event_count
    rules:
        - ts-skip-rule
    group-by:
        - User
    timespan: 10s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig {
            timestamp_fallback: TimestampFallback::Skip,
            ..Default::default()
        });
        engine.add_collection(&collection).unwrap();
        assert_eq!(engine.correlation_rule_count(), 1);

        // Events with no timestamp field — should NOT update correlation state
        let v = json!({"action": "click", "User": "alice"});
        let event = Event::from_value(&v);

        let r1 = engine.process_event(&event);
        assert!(!r1.detections.is_empty(), "detection should still fire");

        let r2 = engine.process_event(&event);
        assert!(!r2.detections.is_empty(), "detection should still fire");

        let r3 = engine.process_event(&event);
        assert!(!r3.detections.is_empty(), "detection should still fire");

        // No correlations should fire because events were skipped
        assert!(r1.correlations.is_empty());
        assert!(r2.correlations.is_empty());
        assert!(r3.correlations.is_empty());
    }

    #[test]
    fn test_timestamp_fallback_wallclock_default() {
        let yaml = r#"
title: test rule
id: ts-wc-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: test correlation
correlation:
    type: event_count
    rules:
        - ts-wc-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();
        assert_eq!(engine.correlation_rule_count(), 1);

        // Events with no timestamp — WallClock fallback means they get Utc::now()
        // and should be close enough to correlate (generous 60s window)
        let v = json!({"action": "click", "User": "alice"});
        let event = Event::from_value(&v);

        let _r1 = engine.process_event(&event);
        let _r2 = engine.process_event(&event);
        let r3 = engine.process_event(&event);

        // With WallClock, all events get near-identical timestamps and should correlate
        assert!(
            !r3.correlations.is_empty(),
            "WallClock fallback should allow correlation"
        );
    }

    // =========================================================================
    // Event count correlation
    // =========================================================================

    #[test]
    fn test_event_count_basic() {
        let yaml = r#"
title: Base Rule
id: base-rule-001
name: base_rule
logsource:
    product: windows
    category: process_creation
detection:
    selection:
        CommandLine|contains: 'whoami'
    condition: selection
level: low
---
title: Multiple Whoami
id: corr-001
correlation:
    type: event_count
    rules:
        - base-rule-001
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        assert_eq!(engine.detection_rule_count(), 1);
        assert_eq!(engine.correlation_rule_count(), 1);

        // Send 3 events from same user within the window
        let base_ts = 1000i64;
        for i in 0..3 {
            let v = json!({"CommandLine": "whoami", "User": "admin"});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, base_ts + i * 10);

            // Each event should match the detection rule
            assert_eq!(result.detections.len(), 1);

            if i < 2 {
                // Not enough events yet
                assert!(result.correlations.is_empty());
            } else {
                // 3rd event triggers the correlation
                assert_eq!(result.correlations.len(), 1);
                assert_eq!(result.correlations[0].rule_title, "Multiple Whoami");
                assert_eq!(result.correlations[0].aggregated_value, 3.0);
            }
        }
    }

    #[test]
    fn test_event_count_different_groups() {
        let yaml = r#"
title: Login
id: login-001
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
level: low
---
title: Many Logins
id: corr-login
correlation:
    type: event_count
    rules:
        - login-001
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // User "alice" sends 2 events, "bob" sends 3
        let ts = 1000i64;
        for i in 0..2 {
            let v = json!({"EventType": "login", "User": "alice"});
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, ts + i);
            assert!(r.correlations.is_empty());
        }
        for i in 0..3 {
            let v = json!({"EventType": "login", "User": "bob"});
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, ts + i);
            if i == 2 {
                assert_eq!(r.correlations.len(), 1);
                assert_eq!(
                    r.correlations[0].group_key,
                    vec![("User".to_string(), "bob".to_string())]
                );
            }
        }
    }

    #[test]
    fn test_event_count_window_expiry() {
        let yaml = r#"
title: Base
id: base-002
logsource:
    category: test
detection:
    selection:
        action: click
    condition: selection
---
title: Rapid Clicks
id: corr-002
correlation:
    type: event_count
    rules:
        - base-002
    group-by:
        - User
    timespan: 10s
    condition:
        gte: 3
level: medium
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // Send 2 events at t=0,1 then 1 event at t=15 (outside window)
        let v = json!({"action": "click", "User": "admin"});
        let event = Event::from_value(&v);
        engine.process_event_at(&event, 0);
        engine.process_event_at(&event, 1);
        let r = engine.process_event_at(&event, 15);
        // Only 1 event in window [5, 15], not enough
        assert!(r.correlations.is_empty());
    }

    // =========================================================================
    // Value count correlation
    // =========================================================================

    #[test]
    fn test_value_count() {
        let yaml = r#"
title: Failed Login
id: failed-login-001
logsource:
    category: auth
detection:
    selection:
        EventType: failed_login
    condition: selection
level: low
---
title: Failed Logins From Many Users
id: corr-vc-001
correlation:
    type: value_count
    rules:
        - failed-login-001
    group-by:
        - Host
    timespan: 60s
    condition:
        field: User
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // 3 different users failing login on same host
        for (i, user) in ["alice", "bob", "charlie"].iter().enumerate() {
            let v = json!({"EventType": "failed_login", "Host": "srv01", "User": user});
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, ts + i as i64);
            if i == 2 {
                assert_eq!(r.correlations.len(), 1);
                assert_eq!(r.correlations[0].aggregated_value, 3.0);
            }
        }
    }

    // =========================================================================
    // Temporal correlation
    // =========================================================================

    #[test]
    fn test_temporal() {
        let yaml = r#"
title: Recon A
id: recon-a
name: recon_a
logsource:
    category: process
detection:
    selection:
        CommandLine|contains: 'whoami'
    condition: selection
---
title: Recon B
id: recon-b
name: recon_b
logsource:
    category: process
detection:
    selection:
        CommandLine|contains: 'ipconfig'
    condition: selection
---
title: Recon Combo
id: corr-temporal
correlation:
    type: temporal
    rules:
        - recon-a
        - recon-b
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Only recon A fires
        let v1 = json!({"CommandLine": "whoami", "User": "admin"});
        let ev1 = Event::from_value(&v1);
        let r1 = engine.process_event_at(&ev1, ts);
        assert!(r1.correlations.is_empty());

        // Now recon B fires — both rules have fired within window
        let v2 = json!({"CommandLine": "ipconfig /all", "User": "admin"});
        let ev2 = Event::from_value(&v2);
        let r2 = engine.process_event_at(&ev2, ts + 10);
        assert_eq!(r2.correlations.len(), 1);
        assert_eq!(r2.correlations[0].rule_title, "Recon Combo");
    }

    // =========================================================================
    // Temporal ordered correlation
    // =========================================================================

    #[test]
    fn test_temporal_ordered() {
        let yaml = r#"
title: Failed Login
id: failed-001
name: failed_login
logsource:
    category: auth
detection:
    selection:
        EventType: failed_login
    condition: selection
---
title: Success Login
id: success-001
name: successful_login
logsource:
    category: auth
detection:
    selection:
        EventType: success_login
    condition: selection
---
title: Brute Force Then Login
id: corr-bf
correlation:
    type: temporal_ordered
    rules:
        - failed-001
        - success-001
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: critical
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Failed login first
        let v1 = json!({"EventType": "failed_login", "User": "admin"});
        let ev1 = Event::from_value(&v1);
        let r1 = engine.process_event_at(&ev1, ts);
        assert!(r1.correlations.is_empty());

        // Then successful login — correct order!
        let v2 = json!({"EventType": "success_login", "User": "admin"});
        let ev2 = Event::from_value(&v2);
        let r2 = engine.process_event_at(&ev2, ts + 10);
        assert_eq!(r2.correlations.len(), 1);
    }

    #[test]
    fn test_temporal_ordered_wrong_order() {
        let yaml = r#"
title: Rule A
id: rule-a
logsource:
    category: test
detection:
    selection:
        type: a
    condition: selection
---
title: Rule B
id: rule-b
logsource:
    category: test
detection:
    selection:
        type: b
    condition: selection
---
title: A then B
id: corr-ab
correlation:
    type: temporal_ordered
    rules:
        - rule-a
        - rule-b
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // B fires first, then A — wrong order
        let v1 = json!({"type": "b", "User": "admin"});
        let ev1 = Event::from_value(&v1);
        engine.process_event_at(&ev1, ts);

        let v2 = json!({"type": "a", "User": "admin"});
        let ev2 = Event::from_value(&v2);
        let r2 = engine.process_event_at(&ev2, ts + 10);
        assert!(r2.correlations.is_empty());
    }

    // =========================================================================
    // Numeric aggregation (value_sum, value_avg)
    // =========================================================================

    #[test]
    fn test_value_sum() {
        let yaml = r#"
title: Web Access
id: web-001
logsource:
    category: web
detection:
    selection:
        action: upload
    condition: selection
---
title: Large Upload
id: corr-sum
correlation:
    type: value_sum
    rules:
        - web-001
    group-by:
        - User
    timespan: 60s
    condition:
        field: bytes_sent
        gt: 1000
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        let v1 = json!({"action": "upload", "User": "alice", "bytes_sent": 600});
        let ev1 = Event::from_value(&v1);
        let r1 = engine.process_event_at(&ev1, ts);
        assert!(r1.correlations.is_empty());

        let v2 = json!({"action": "upload", "User": "alice", "bytes_sent": 500});
        let ev2 = Event::from_value(&v2);
        let r2 = engine.process_event_at(&ev2, ts + 5);
        assert_eq!(r2.correlations.len(), 1);
        assert!((r2.correlations[0].aggregated_value - 1100.0).abs() < f64::EPSILON);
    }

    #[test]
    fn test_value_avg() {
        let yaml = r#"
title: Request
id: req-001
logsource:
    category: web
detection:
    selection:
        type: request
    condition: selection
---
title: High Avg Latency
id: corr-avg
correlation:
    type: value_avg
    rules:
        - req-001
    group-by:
        - Service
    timespan: 60s
    condition:
        field: latency_ms
        gt: 500
level: medium
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Avg of 400, 600, 800 = 600 > 500
        for (i, latency) in [400, 600, 800].iter().enumerate() {
            let v = json!({"type": "request", "Service": "api", "latency_ms": latency});
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, ts + i as i64);
            if i == 2 {
                assert_eq!(r.correlations.len(), 1);
                assert!((r.correlations[0].aggregated_value - 600.0).abs() < f64::EPSILON);
            }
        }
    }

    // =========================================================================
    // State management
    // =========================================================================

    #[test]
    fn test_state_count() {
        let yaml = r#"
title: Base
id: base-sc
logsource:
    category: test
detection:
    selection:
        action: test
    condition: selection
---
title: Count
id: corr-sc
correlation:
    type: event_count
    rules:
        - base-sc
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 100
level: low
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let v = json!({"action": "test", "User": "alice"});
        let event = Event::from_value(&v);
        engine.process_event_at(&event, 1000);
        assert_eq!(engine.state_count(), 1);

        let v2 = json!({"action": "test", "User": "bob"});
        let event2 = Event::from_value(&v2);
        engine.process_event_at(&event2, 1001);
        assert_eq!(engine.state_count(), 2);

        // Evict everything
        engine.evict_expired(2000);
        assert_eq!(engine.state_count(), 0);
    }

    // =========================================================================
    // Generate flag
    // =========================================================================

    #[test]
    fn test_generate_flag_default_false() {
        let yaml = r#"
title: Base
id: gen-base
logsource:
    category: test
detection:
    selection:
        action: test
    condition: selection
---
title: Correlation
id: gen-corr
correlation:
    type: event_count
    rules:
        - gen-base
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 1
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // generate defaults to false — detection matches are still returned
        // (filtering by generate flag is a backend concern, not eval)
        let v = json!({"action": "test", "User": "alice"});
        let event = Event::from_value(&v);
        let r = engine.process_event_at(&event, 1000);
        assert_eq!(r.detections.len(), 1);
        assert_eq!(r.correlations.len(), 1);
    }

    // =========================================================================
    // Real-world example: AWS bucket enumeration
    // =========================================================================

    #[test]
    fn test_aws_bucket_enumeration() {
        let yaml = r#"
title: Potential Bucket Enumeration on AWS
id: f305fd62-beca-47da-ad95-7690a0620084
logsource:
    product: aws
    service: cloudtrail
detection:
    selection:
        eventSource: "s3.amazonaws.com"
        eventName: "ListBuckets"
    condition: selection
level: low
---
title: Multiple AWS bucket enumerations
id: be246094-01d3-4bba-88de-69e582eba0cc
status: experimental
correlation:
    type: event_count
    rules:
        - f305fd62-beca-47da-ad95-7690a0620084
    group-by:
        - userIdentity.arn
    timespan: 1h
    condition:
        gte: 5
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let base_ts = 1_700_000_000i64;
        for i in 0..5 {
            let v = json!({
                "eventSource": "s3.amazonaws.com",
                "eventName": "ListBuckets",
                "userIdentity.arn": "arn:aws:iam::123456789:user/attacker"
            });
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, base_ts + i * 60);
            if i == 4 {
                assert_eq!(r.correlations.len(), 1);
                assert_eq!(
                    r.correlations[0].rule_title,
                    "Multiple AWS bucket enumerations"
                );
                assert_eq!(r.correlations[0].aggregated_value, 5.0);
            }
        }
    }

    // =========================================================================
    // Chaining: event_count -> temporal_ordered
    // =========================================================================

    #[test]
    fn test_chaining_event_count_to_temporal() {
        // Reproduces the spec's "failed logins followed by successful login" example.
        // Chain: failed_login (detection) -> many_failed (event_count) -> brute_then_login (temporal_ordered)
        let yaml = r#"
title: Single failed login
id: failed-login-chain
name: failed_login
logsource:
    category: auth
detection:
    selection:
        EventType: failed_login
    condition: selection
---
title: Successful login
id: success-login-chain
name: successful_login
logsource:
    category: auth
detection:
    selection:
        EventType: success_login
    condition: selection
---
title: Multiple failed logins
id: many-failed-chain
name: multiple_failed_login
correlation:
    type: event_count
    rules:
        - failed-login-chain
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: medium
---
title: Brute Force Followed by Login
id: brute-force-chain
correlation:
    type: temporal_ordered
    rules:
        - many-failed-chain
        - success-login-chain
    group-by:
        - User
    timespan: 120s
    condition:
        gte: 2
level: critical
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        assert_eq!(engine.detection_rule_count(), 2);
        assert_eq!(engine.correlation_rule_count(), 2);

        let ts = 1000i64;

        // Send 3 failed logins → triggers "many_failed_chain"
        for i in 0..3 {
            let v = json!({"EventType": "failed_login", "User": "victim"});
            let event = Event::from_value(&v);
            let r = engine.process_event_at(&event, ts + i);
            if i == 2 {
                // The event_count correlation should fire
                assert!(
                    r.correlations
                        .iter()
                        .any(|c| c.rule_title == "Multiple failed logins"),
                    "Expected event_count correlation to fire"
                );
            }
        }

        // Now send a successful login → should trigger the chained temporal_ordered
        // Note: chaining happens in chain_correlations when many-failed-chain fires
        // and then success-login-chain matches the detection.
        // The temporal_ordered correlation needs BOTH many-failed-chain AND success-login-chain
        // to have fired. success-login-chain is a detection rule, not a correlation,
        // so it gets matched via the regular detection path.
        let v = json!({"EventType": "success_login", "User": "victim"});
        let event = Event::from_value(&v);
        let r = engine.process_event_at(&event, ts + 30);

        // The detection should match
        assert_eq!(r.detections.len(), 1);
        assert_eq!(r.detections[0].rule_title, "Successful login");
    }

    // =========================================================================
    // Field aliases
    // =========================================================================

    #[test]
    fn test_field_aliases() {
        let yaml = r#"
title: Internal Error
id: internal-error-001
name: internal_error
logsource:
    category: web
detection:
    selection:
        http.response.status_code: 500
    condition: selection
---
title: New Connection
id: new-conn-001
name: new_network_connection
logsource:
    category: network
detection:
    selection:
        event.type: connection
    condition: selection
---
title: Error Then Connection
id: corr-alias
correlation:
    type: temporal
    rules:
        - internal-error-001
        - new-conn-001
    group-by:
        - internal_ip
    timespan: 60s
    condition:
        gte: 2
    aliases:
        internal_ip:
            internal_error: destination.ip
            new_network_connection: source.ip
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;

        // Internal error with destination.ip = 10.0.0.5
        let v1 = json!({
            "http.response.status_code": 500,
            "destination.ip": "10.0.0.5"
        });
        let ev1 = Event::from_value(&v1);
        let r1 = engine.process_event_at(&ev1, ts);
        assert_eq!(r1.detections.len(), 1);
        assert!(r1.correlations.is_empty());

        // New connection with source.ip = 10.0.0.5 (same IP, aliased)
        let v2 = json!({
            "event.type": "connection",
            "source.ip": "10.0.0.5"
        });
        let ev2 = Event::from_value(&v2);
        let r2 = engine.process_event_at(&ev2, ts + 5);
        assert_eq!(r2.detections.len(), 1);
        // Both rules fired for the same internal_ip group → temporal should fire
        assert_eq!(r2.correlations.len(), 1);
        assert_eq!(r2.correlations[0].rule_title, "Error Then Connection");
        // Check group key contains the aliased field
        assert!(
            r2.correlations[0]
                .group_key
                .iter()
                .any(|(k, v)| k == "internal_ip" && v == "10.0.0.5")
        );
    }

    // =========================================================================
    // Value percentile (basic smoke test)
    // =========================================================================

    #[test]
    fn test_value_percentile() {
        let yaml = r#"
title: Process Creation
id: proc-001
logsource:
    category: process
detection:
    selection:
        type: process_creation
    condition: selection
---
title: Rare Process
id: corr-percentile
correlation:
    type: value_percentile
    rules:
        - proc-001
    group-by:
        - ComputerName
    timespan: 60s
    condition:
        field: image
        lte: 50
level: medium
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Push some numeric-ish values for the image field
        for (i, val) in [10.0, 20.0, 30.0, 40.0, 50.0].iter().enumerate() {
            let v = json!({"type": "process_creation", "ComputerName": "srv01", "image": val});
            let event = Event::from_value(&v);
            let _ = engine.process_event_at(&event, ts + i as i64);
        }
        // The median (30.0) should be <= 50, so condition fires
        // Note: percentile implementation is simplified for in-memory eval
    }

    // =========================================================================
    // Extended temporal conditions (end-to-end)
    // =========================================================================

    #[test]
    fn test_extended_temporal_and_condition() {
        // Temporal correlation with "rule_a and rule_b" extended condition
        let yaml = r#"
title: Login Attempt
id: login-attempt
logsource:
    category: auth
detection:
    selection:
        EventType: login_failure
    condition: selection
---
title: Password Change
id: password-change
logsource:
    category: auth
detection:
    selection:
        EventType: password_change
    condition: selection
---
title: Credential Attack
correlation:
    type: temporal
    rules:
        - login-attempt
        - password-change
    group-by:
        - User
    timespan: 300s
    condition: login-attempt and password-change
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;

        // Login failure by alice
        let ev1 = json!({"EventType": "login_failure", "User": "alice"});
        let r1 = engine.process_event_at(&Event::from_value(&ev1), ts);
        assert!(r1.correlations.is_empty(), "only one rule fired so far");

        // Password change by alice — both rules have now fired
        let ev2 = json!({"EventType": "password_change", "User": "alice"});
        let r2 = engine.process_event_at(&Event::from_value(&ev2), ts + 10);
        assert_eq!(
            r2.correlations.len(),
            1,
            "temporal correlation should fire: both rules matched"
        );
        assert_eq!(r2.correlations[0].rule_title, "Credential Attack");
    }

    #[test]
    fn test_extended_temporal_or_condition() {
        // Temporal with "rule_a or rule_b" — should fire when either fires
        let yaml = r#"
title: SSH Login
id: ssh-login
logsource:
    category: auth
detection:
    selection:
        EventType: ssh_login
    condition: selection
---
title: VPN Login
id: vpn-login
logsource:
    category: auth
detection:
    selection:
        EventType: vpn_login
    condition: selection
---
title: Any Remote Access
correlation:
    type: temporal
    rules:
        - ssh-login
        - vpn-login
    group-by:
        - User
    timespan: 60s
    condition: ssh-login or vpn-login
level: medium
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // Only SSH login by bob — "or" means this suffices
        let ev = json!({"EventType": "ssh_login", "User": "bob"});
        let r = engine.process_event_at(&Event::from_value(&ev), 1000);
        assert_eq!(r.correlations.len(), 1);
        assert_eq!(r.correlations[0].rule_title, "Any Remote Access");
    }

    #[test]
    fn test_extended_temporal_partial_and_no_fire() {
        // Temporal "and" with only one rule firing should not trigger
        let yaml = r#"
title: Recon Step 1
id: recon-1
logsource:
    category: process
detection:
    selection:
        CommandLine|contains: 'whoami'
    condition: selection
---
title: Recon Step 2
id: recon-2
logsource:
    category: process
detection:
    selection:
        CommandLine|contains: 'ipconfig'
    condition: selection
---
title: Full Recon
correlation:
    type: temporal
    rules:
        - recon-1
        - recon-2
    group-by:
        - Host
    timespan: 120s
    condition: recon-1 and recon-2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // Only whoami (recon-1) — should not fire
        let ev = json!({"CommandLine": "whoami", "Host": "srv01"});
        let r = engine.process_event_at(&Event::from_value(&ev), 1000);
        assert!(r.correlations.is_empty(), "only one of two AND rules fired");

        // Now ipconfig (recon-2) — should fire
        let ev2 = json!({"CommandLine": "ipconfig /all", "Host": "srv01"});
        let r2 = engine.process_event_at(&Event::from_value(&ev2), 1010);
        assert_eq!(r2.correlations.len(), 1);
        assert_eq!(r2.correlations[0].rule_title, "Full Recon");
    }

    // =========================================================================
    // Filter rules with correlation engine
    // =========================================================================

    #[test]
    fn test_filter_with_correlation() {
        // Detection rule + filter + event_count correlation
        let yaml = r#"
title: Failed Auth
id: failed-auth
logsource:
    category: auth
detection:
    selection:
        EventType: auth_failure
    condition: selection
---
title: Exclude Service Accounts
filter:
    rules:
        - failed-auth
    selection:
        User|startswith: 'svc_'
    condition: selection
---
title: Brute Force
correlation:
    type: event_count
    rules:
        - failed-auth
    group-by:
        - User
    timespan: 300s
    condition:
        gte: 3
level: critical
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;

        // Service account failures should be filtered — don't count
        for i in 0..5 {
            let ev = json!({"EventType": "auth_failure", "User": "svc_backup"});
            let r = engine.process_event_at(&Event::from_value(&ev), ts + i);
            assert!(
                r.correlations.is_empty(),
                "service account should be filtered, no correlation"
            );
        }

        // Normal user failures should count
        for i in 0..2 {
            let ev = json!({"EventType": "auth_failure", "User": "alice"});
            let r = engine.process_event_at(&Event::from_value(&ev), ts + 10 + i);
            assert!(r.correlations.is_empty(), "not yet 3 events");
        }

        // Third failure triggers correlation
        let ev = json!({"EventType": "auth_failure", "User": "alice"});
        let r = engine.process_event_at(&Event::from_value(&ev), ts + 12);
        assert_eq!(r.correlations.len(), 1);
        assert_eq!(r.correlations[0].rule_title, "Brute Force");
    }

    // =========================================================================
    // action: repeat with correlation engine
    // =========================================================================

    #[test]
    fn test_repeat_rules_in_correlation() {
        // Two detection rules via repeat, both feed into event_count
        let yaml = r#"
title: File Access A
id: file-a
logsource:
    category: file_access
detection:
    selection:
        FileName|endswith: '.docx'
    condition: selection
---
action: repeat
title: File Access B
id: file-b
detection:
    selection:
        FileName|endswith: '.xlsx'
    condition: selection
---
title: Mass File Access
correlation:
    type: event_count
    rules:
        - file-a
        - file-b
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        assert_eq!(collection.rules.len(), 2);
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();
        assert_eq!(engine.detection_rule_count(), 2);

        let ts = 1000i64;
        // Mix of docx and xlsx accesses by same user
        let ev1 = json!({"FileName": "report.docx", "User": "bob"});
        engine.process_event_at(&Event::from_value(&ev1), ts);
        let ev2 = json!({"FileName": "data.xlsx", "User": "bob"});
        engine.process_event_at(&Event::from_value(&ev2), ts + 1);
        let ev3 = json!({"FileName": "notes.docx", "User": "bob"});
        let r = engine.process_event_at(&Event::from_value(&ev3), ts + 2);

        assert_eq!(r.correlations.len(), 1);
        assert_eq!(r.correlations[0].rule_title, "Mass File Access");
    }

    // =========================================================================
    // Expand modifier with correlation engine
    // =========================================================================

    #[test]
    fn test_expand_modifier_with_correlation() {
        let yaml = r#"
title: User Temp File
id: user-temp
logsource:
    category: file_access
detection:
    selection:
        FilePath|expand: 'C:\Users\%User%\Temp'
    condition: selection
---
title: Excessive Temp Access
correlation:
    type: event_count
    rules:
        - user-temp
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: medium
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Event where User field matches the placeholder
        let ev1 = json!({"FilePath": "C:\\Users\\alice\\Temp", "User": "alice"});
        let r1 = engine.process_event_at(&Event::from_value(&ev1), ts);
        assert!(r1.correlations.is_empty());

        let ev2 = json!({"FilePath": "C:\\Users\\alice\\Temp", "User": "alice"});
        let r2 = engine.process_event_at(&Event::from_value(&ev2), ts + 1);
        assert_eq!(r2.correlations.len(), 1);
        assert_eq!(r2.correlations[0].rule_title, "Excessive Temp Access");

        // Different user — should NOT match (path says alice, user is bob)
        let ev3 = json!({"FilePath": "C:\\Users\\alice\\Temp", "User": "bob"});
        let r3 = engine.process_event_at(&Event::from_value(&ev3), ts + 2);
        // Detection doesn't fire for this event since expand resolves to C:\Users\bob\Temp
        assert_eq!(r3.detections.len(), 0);
    }

    // =========================================================================
    // Timestamp modifier with correlation engine
    // =========================================================================

    #[test]
    fn test_timestamp_modifier_with_correlation() {
        let yaml = r#"
title: Night Login
id: night-login
logsource:
    category: auth
detection:
    login:
        EventType: login
    night:
        Timestamp|hour: 3
    condition: login and night
---
title: Frequent Night Logins
correlation:
    type: event_count
    rules:
        - night-login
    group-by:
        - User
    timespan: 3600s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts = 1000i64;
        // Login at 3AM
        let ev1 =
            json!({"EventType": "login", "User": "alice", "Timestamp": "2024-01-15T03:10:00Z"});
        let r1 = engine.process_event_at(&Event::from_value(&ev1), ts);
        assert_eq!(r1.detections.len(), 1);
        assert!(r1.correlations.is_empty());

        let ev2 =
            json!({"EventType": "login", "User": "alice", "Timestamp": "2024-01-15T03:45:00Z"});
        let r2 = engine.process_event_at(&Event::from_value(&ev2), ts + 1);
        assert_eq!(r2.correlations.len(), 1);
        assert_eq!(r2.correlations[0].rule_title, "Frequent Night Logins");

        // Login at noon — should NOT count
        let ev3 = json!({"EventType": "login", "User": "bob", "Timestamp": "2024-01-15T12:00:00Z"});
        let r3 = engine.process_event_at(&Event::from_value(&ev3), ts + 2);
        assert!(
            r3.detections.is_empty(),
            "noon login should not match night rule"
        );
    }

    // =========================================================================
    // Correlation condition range (multiple predicates)
    // =========================================================================

    #[test]
    fn test_event_count_range_condition() {
        let yaml = r#"
title: Login Attempt
id: login-attempt-001
name: login_attempt
logsource:
    product: windows
detection:
    selection:
        EventType: login
    condition: selection
level: low
---
title: Login Count Range
id: corr-range-001
correlation:
    type: event_count
    rules:
        - login-attempt-001
    group-by:
        - User
    timespan: 3600s
    condition:
        gt: 2
        lte: 5
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ts: i64 = 1_000_000;

        // Send 2 events — gt:2 is false
        for i in 0..2 {
            let ev = json!({"EventType": "login", "User": "alice"});
            let r = engine.process_event_at(&Event::from_value(&ev), ts + i);
            assert!(r.correlations.is_empty(), "2 events should not fire (gt:2)");
        }

        // 3rd event — gt:2 is true, lte:5 is true → fires
        let ev3 = json!({"EventType": "login", "User": "alice"});
        let r3 = engine.process_event_at(&Event::from_value(&ev3), ts + 3);
        assert_eq!(r3.correlations.len(), 1, "3 events: gt:2 AND lte:5");

        // Send events 4, 5 — still in range
        for i in 4..=5 {
            let ev = json!({"EventType": "login", "User": "alice"});
            let r = engine.process_event_at(&Event::from_value(&ev), ts + i);
            assert_eq!(r.correlations.len(), 1, "{i} events still in range");
        }

        // 6th event — lte:5 is false → no fire
        let ev6 = json!({"EventType": "login", "User": "alice"});
        let r6 = engine.process_event_at(&Event::from_value(&ev6), ts + 6);
        assert!(
            r6.correlations.is_empty(),
            "6 events exceeds lte:5, should not fire"
        );
    }

    // =========================================================================
    // Suppression
    // =========================================================================

    fn suppression_yaml() -> &'static str {
        r#"
title: Login
id: login-base
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-base
    group-by:
        - User
    timeframe: 60s
    condition:
        gte: 3
level: high
"#
    }

    #[test]
    fn test_suppression_window() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let config = CorrelationConfig {
            suppress: Some(10), // suppress for 10 seconds
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let ts = 1000;

        // Fire 3 events to hit threshold
        engine.process_event_at(&Event::from_value(&ev), ts);
        engine.process_event_at(&Event::from_value(&ev), ts + 1);
        let r3 = engine.process_event_at(&Event::from_value(&ev), ts + 2);
        assert_eq!(r3.correlations.len(), 1, "should fire on 3rd event");

        // 4th event within suppress window → suppressed
        let r4 = engine.process_event_at(&Event::from_value(&ev), ts + 3);
        assert!(
            r4.correlations.is_empty(),
            "should be suppressed within 10s window"
        );

        // 5th event still within suppress window → suppressed
        let r5 = engine.process_event_at(&Event::from_value(&ev), ts + 9);
        assert!(
            r5.correlations.is_empty(),
            "should be suppressed at ts+9 (< ts+2+10)"
        );

        // Event after suppress window expires → fires again
        let r6 = engine.process_event_at(&Event::from_value(&ev), ts + 13);
        assert_eq!(
            r6.correlations.len(),
            1,
            "should fire again after suppress window expires"
        );
    }

    #[test]
    fn test_suppression_per_group_key() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let config = CorrelationConfig {
            suppress: Some(60),
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ts = 1000;

        // Alice hits threshold
        let ev_a = json!({"EventType": "login", "User": "alice"});
        engine.process_event_at(&Event::from_value(&ev_a), ts);
        engine.process_event_at(&Event::from_value(&ev_a), ts + 1);
        let r = engine.process_event_at(&Event::from_value(&ev_a), ts + 2);
        assert_eq!(r.correlations.len(), 1, "alice should fire");

        // Bob hits threshold — different group key, not suppressed
        let ev_b = json!({"EventType": "login", "User": "bob"});
        engine.process_event_at(&Event::from_value(&ev_b), ts + 3);
        engine.process_event_at(&Event::from_value(&ev_b), ts + 4);
        let r = engine.process_event_at(&Event::from_value(&ev_b), ts + 5);
        assert_eq!(r.correlations.len(), 1, "bob should fire independently");

        // Alice is still suppressed
        let r = engine.process_event_at(&Event::from_value(&ev_a), ts + 6);
        assert!(r.correlations.is_empty(), "alice still suppressed");
    }

    // =========================================================================
    // Action on match: Reset
    // =========================================================================

    #[test]
    fn test_action_reset() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let config = CorrelationConfig {
            action_on_match: CorrelationAction::Reset,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let ts = 1000;

        // Hit threshold: 3 events
        engine.process_event_at(&Event::from_value(&ev), ts);
        engine.process_event_at(&Event::from_value(&ev), ts + 1);
        let r3 = engine.process_event_at(&Event::from_value(&ev), ts + 2);
        assert_eq!(r3.correlations.len(), 1, "should fire on 3rd event");

        // State was reset, so 4th and 5th events should NOT fire
        let r4 = engine.process_event_at(&Event::from_value(&ev), ts + 3);
        assert!(r4.correlations.is_empty(), "reset: need 3 more events");

        let r5 = engine.process_event_at(&Event::from_value(&ev), ts + 4);
        assert!(r5.correlations.is_empty(), "reset: still only 2");

        // 6th event (3rd after reset) should fire again
        let r6 = engine.process_event_at(&Event::from_value(&ev), ts + 5);
        assert_eq!(
            r6.correlations.len(),
            1,
            "should fire again after 3 events post-reset"
        );
    }

    // =========================================================================
    // Generate flag / emit_detections
    // =========================================================================

    #[test]
    fn test_emit_detections_true_by_default() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let r = engine.process_event_at(&Event::from_value(&ev), 1000);
        assert_eq!(r.detections.len(), 1, "by default detections are emitted");
    }

    #[test]
    fn test_emit_detections_false_suppresses() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let config = CorrelationConfig {
            emit_detections: false,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let r = engine.process_event_at(&Event::from_value(&ev), 1000);
        assert!(
            r.detections.is_empty(),
            "detection matches should be suppressed when emit_detections=false"
        );
    }

    #[test]
    fn test_generate_true_keeps_detections() {
        // When generate: true, detections should be emitted even with emit_detections=false
        let yaml = r#"
title: Login
id: login-gen
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-gen
    group-by:
        - User
    timeframe: 60s
    condition:
        gte: 3
    generate: true
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            emit_detections: false,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let r = engine.process_event_at(&Event::from_value(&ev), 1000);
        // generate: true means this rule is NOT correlation-only
        assert_eq!(
            r.detections.len(),
            1,
            "generate:true keeps detection output"
        );
    }

    // =========================================================================
    // Suppression + Reset combined
    // =========================================================================

    #[test]
    fn test_suppress_and_reset_combined() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let config = CorrelationConfig {
            suppress: Some(5),
            action_on_match: CorrelationAction::Reset,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let ts = 1000;

        // Hit threshold: fires and resets
        engine.process_event_at(&Event::from_value(&ev), ts);
        engine.process_event_at(&Event::from_value(&ev), ts + 1);
        let r3 = engine.process_event_at(&Event::from_value(&ev), ts + 2);
        assert_eq!(r3.correlations.len(), 1, "fires on 3rd event");

        // Push 3 more events quickly (state was reset, so new count → 3)
        // but suppress window hasn't expired (ts+2 + 5 = ts+7)
        engine.process_event_at(&Event::from_value(&ev), ts + 3);
        engine.process_event_at(&Event::from_value(&ev), ts + 4);
        let r = engine.process_event_at(&Event::from_value(&ev), ts + 5);
        assert!(
            r.correlations.is_empty(),
            "threshold met again but still suppressed"
        );

        // After suppress expires (at ts+8, which is ts+2+6 > suppress=5),
        // the accumulated events from step 2 (ts+3,4,5) still satisfy gte:3,
        // so the first event after expiry fires immediately and resets.
        let r = engine.process_event_at(&Event::from_value(&ev), ts + 8);
        assert_eq!(
            r.correlations.len(),
            1,
            "fires after suppress expires (accumulated events + new one)"
        );

        // State was reset again at ts+8, suppress window now ts+8..ts+13.
        // Need 3 new events to fire, and suppress must expire.
        engine.process_event_at(&Event::from_value(&ev), ts + 9);
        engine.process_event_at(&Event::from_value(&ev), ts + 10);
        let r = engine.process_event_at(&Event::from_value(&ev), ts + 11);
        assert!(
            r.correlations.is_empty(),
            "threshold met but suppress window hasn't expired (ts+11 - ts+8 = 3 < 5)"
        );
    }

    // =========================================================================
    // No suppression (default behavior preserved)
    // =========================================================================

    #[test]
    fn test_no_suppression_fires_every_event() {
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        let ev = json!({"EventType": "login", "User": "alice"});
        let ts = 1000;

        engine.process_event_at(&Event::from_value(&ev), ts);
        engine.process_event_at(&Event::from_value(&ev), ts + 1);
        let r3 = engine.process_event_at(&Event::from_value(&ev), ts + 2);
        assert_eq!(r3.correlations.len(), 1);

        // Without suppression, 4th event should also fire
        let r4 = engine.process_event_at(&Event::from_value(&ev), ts + 3);
        assert_eq!(
            r4.correlations.len(),
            1,
            "no suppression: fires on every event after threshold"
        );

        let r5 = engine.process_event_at(&Event::from_value(&ev), ts + 4);
        assert_eq!(r5.correlations.len(), 1, "still fires");
    }

    // =========================================================================
    // Custom attribute → engine config tests
    // =========================================================================

    #[test]
    fn test_custom_attr_timestamp_field() {
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.timestamp_field".to_string(), "time".to_string());
        engine.apply_custom_attributes(&attrs);

        assert_eq!(
            engine.config.timestamp_fields[0], "time",
            "rsigma.timestamp_field should be prepended"
        );
        // Defaults should still be there after the custom one
        assert!(
            engine
                .config
                .timestamp_fields
                .contains(&"@timestamp".to_string())
        );
    }

    #[test]
    fn test_custom_attr_timestamp_field_no_duplicates() {
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.timestamp_field".to_string(), "time".to_string());
        // Apply twice — should not duplicate
        engine.apply_custom_attributes(&attrs);
        engine.apply_custom_attributes(&attrs);

        let count = engine
            .config
            .timestamp_fields
            .iter()
            .filter(|f| *f == "time")
            .count();
        assert_eq!(count, 1, "should not duplicate timestamp_field entries");
    }

    #[test]
    fn test_custom_attr_suppress() {
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        assert!(engine.config.suppress.is_none());

        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.suppress".to_string(), "5m".to_string());
        engine.apply_custom_attributes(&attrs);

        assert_eq!(engine.config.suppress, Some(300));
    }

    #[test]
    fn test_custom_attr_suppress_does_not_override_cli() {
        let config = CorrelationConfig {
            suppress: Some(60), // CLI set to 60s
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);

        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.suppress".to_string(), "5m".to_string());
        engine.apply_custom_attributes(&attrs);

        assert_eq!(
            engine.config.suppress,
            Some(60),
            "CLI suppress should not be overridden by custom attribute"
        );
    }

    #[test]
    fn test_custom_attr_action() {
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        assert_eq!(engine.config.action_on_match, CorrelationAction::Alert);

        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.action".to_string(), "reset".to_string());
        engine.apply_custom_attributes(&attrs);

        assert_eq!(engine.config.action_on_match, CorrelationAction::Reset);
    }

    #[test]
    fn test_custom_attr_action_does_not_override_cli() {
        let config = CorrelationConfig {
            action_on_match: CorrelationAction::Reset, // CLI set to reset
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);

        let mut attrs = std::collections::HashMap::new();
        attrs.insert("rsigma.action".to_string(), "alert".to_string());
        engine.apply_custom_attributes(&attrs);

        assert_eq!(
            engine.config.action_on_match,
            CorrelationAction::Reset,
            "CLI action should not be overridden by custom attribute"
        );
    }

    #[test]
    fn test_custom_attr_timestamp_field_used_for_extraction() {
        // The event has "time" but not "@timestamp" or "timestamp"
        let collection = parse_sigma_yaml(suppression_yaml()).unwrap();
        let mut config = CorrelationConfig::default();
        // Prepend "event_time" to simulate --timestamp-field
        config.timestamp_fields.insert(0, "event_time".to_string());
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        // Event with "event_time" field
        let ev = json!({
            "EventType": "login",
            "User": "alice",
            "event_time": "2026-02-11T12:00:00Z"
        });
        let result = engine.process_event(&Event::from_value(&ev));

        // The detection should match, and timestamp should be ~1739275200 (2026-02-11)
        assert!(!result.detections.is_empty() || result.correlations.is_empty());
        // The key test: ensure the engine extracted the event timestamp, not Utc::now.
        // If it used Utc::now, the test would still pass but the timestamp would be
        // wildly different. We verify by checking the extracted value directly.
        let ts = engine
            .extract_event_timestamp(&Event::from_value(&ev))
            .expect("should extract timestamp");
        assert!(
            ts > 1_700_000_000 && ts < 1_800_000_000,
            "timestamp should be ~2026 epoch, got {ts}"
        );
    }

    // =========================================================================
    // Cycle detection
    // =========================================================================

    #[test]
    fn test_correlation_cycle_direct() {
        // Two correlations that reference each other: A -> B -> A
        let yaml = r#"
title: detection rule
id: det-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: correlation A
id: corr-a
correlation:
    type: event_count
    rules:
        - corr-b
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
---
title: correlation B
id: corr-b
correlation:
    type: event_count
    rules:
        - corr-a
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let result = engine.add_collection(&collection);
        assert!(result.is_err(), "should detect direct cycle");
        let err = result.unwrap_err().to_string();
        assert!(err.contains("cycle"), "error should mention cycle: {err}");
        assert!(
            err.contains("corr-a") && err.contains("corr-b"),
            "error should name both correlations: {err}"
        );
    }

    #[test]
    fn test_correlation_cycle_self() {
        // A correlation that references itself
        let yaml = r#"
title: detection rule
id: det-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: self-ref correlation
id: self-corr
correlation:
    type: event_count
    rules:
        - self-corr
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let result = engine.add_collection(&collection);
        assert!(result.is_err(), "should detect self-referencing cycle");
        let err = result.unwrap_err().to_string();
        assert!(err.contains("cycle"), "error should mention cycle: {err}");
        assert!(
            err.contains("self-corr"),
            "error should name the correlation: {err}"
        );
    }

    #[test]
    fn test_correlation_no_cycle_valid_chain() {
        // Valid chain: detection -> corr-A -> corr-B (no cycle)
        let yaml = r#"
title: detection rule
id: det-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: correlation A
id: corr-a
correlation:
    type: event_count
    rules:
        - det-rule
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
---
title: correlation B
id: corr-b
correlation:
    type: event_count
    rules:
        - corr-a
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let result = engine.add_collection(&collection);
        assert!(
            result.is_ok(),
            "valid chain should not be rejected: {result:?}"
        );
    }

    #[test]
    fn test_correlation_cycle_transitive() {
        // Transitive cycle: A -> B -> C -> A
        let yaml = r#"
title: detection rule
id: det-rule
logsource:
    product: test
detection:
    selection:
        action: click
    condition: selection
level: low
---
title: correlation A
id: corr-a
correlation:
    type: event_count
    rules:
        - corr-c
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
---
title: correlation B
id: corr-b
correlation:
    type: event_count
    rules:
        - corr-a
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
---
title: correlation C
id: corr-c
correlation:
    type: event_count
    rules:
        - corr-b
    group-by:
        - User
    timespan: 5m
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        let result = engine.add_collection(&collection);
        assert!(result.is_err(), "should detect transitive cycle");
        let err = result.unwrap_err().to_string();
        assert!(err.contains("cycle"), "error should mention cycle: {err}");
    }

    // =========================================================================
    // Correlation event inclusion tests
    // =========================================================================

    #[test]
    fn test_correlation_events_disabled_by_default() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        for i in 0..3 {
            let v = json!({"EventType": "login", "User": "admin", "@timestamp": 1000 + i});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if i == 2 {
                assert_eq!(result.correlations.len(), 1);
                // Events should NOT be included by default
                assert!(result.correlations[0].events.is_none());
            }
        }
    }

    #[test]
    fn test_correlation_events_included_when_enabled() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Full,
            max_correlation_events: 10,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let events_sent: Vec<serde_json::Value> = (0..3)
            .map(|i| json!({"EventType": "login", "User": "admin", "@timestamp": 1000 + i}))
            .collect();

        let mut corr_result = None;
        for (i, ev) in events_sent.iter().enumerate() {
            let event = Event::from_value(ev);
            let result = engine.process_event_at(&event, 1000 + i as i64);
            if !result.correlations.is_empty() {
                corr_result = Some(result);
            }
        }

        let result = corr_result.expect("correlation should have fired");
        let corr = &result.correlations[0];

        // Events should be included
        let events = corr.events.as_ref().expect("events should be present");
        assert_eq!(
            events.len(),
            3,
            "all 3 contributing events should be stored"
        );

        // Verify all sent events are present
        for (i, event) in events.iter().enumerate() {
            assert_eq!(event["EventType"], "login");
            assert_eq!(event["User"], "admin");
            assert_eq!(event["@timestamp"], 1000 + i as i64);
        }
    }

    #[test]
    fn test_correlation_events_max_cap() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 5
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Full,
            max_correlation_events: 3, // only keep last 3
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let mut corr_result = None;
        for i in 0..5 {
            let v = json!({"EventType": "login", "User": "admin", "idx": i});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if !result.correlations.is_empty() {
                corr_result = Some(result);
            }
        }

        let result = corr_result.expect("correlation should have fired");
        let events = result.correlations[0]
            .events
            .as_ref()
            .expect("events should be present");

        // Only the last 3 events should be retained (cap = 3)
        assert_eq!(events.len(), 3);
        assert_eq!(events[0]["idx"], 2);
        assert_eq!(events[1]["idx"], 3);
        assert_eq!(events[2]["idx"], 4);
    }

    #[test]
    fn test_correlation_events_with_reset_action() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Full,
            action_on_match: CorrelationAction::Reset,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        // First round: 2 events -> fires
        for i in 0..2 {
            let v = json!({"EventType": "login", "User": "admin", "round": 1, "idx": i});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if i == 1 {
                assert_eq!(result.correlations.len(), 1);
                let events = result.correlations[0].events.as_ref().unwrap();
                assert_eq!(events.len(), 2);
            }
        }

        // After reset, event buffer should be cleared.
        // Second round: need 2 more events to fire again
        let v = json!({"EventType": "login", "User": "admin", "round": 2, "idx": 0});
        let event = Event::from_value(&v);
        let result = engine.process_event_at(&event, 1010);
        assert!(
            result.correlations.is_empty(),
            "should not fire with only 1 event after reset"
        );

        let v = json!({"EventType": "login", "User": "admin", "round": 2, "idx": 1});
        let event = Event::from_value(&v);
        let result = engine.process_event_at(&event, 1011);
        assert_eq!(result.correlations.len(), 1);
        let events = result.correlations[0].events.as_ref().unwrap();
        assert_eq!(events.len(), 2);
        // Should only have round 2 events
        assert_eq!(events[0]["round"], 2);
        assert_eq!(events[1]["round"], 2);
    }

    #[test]
    fn test_correlation_events_with_set_include() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // Enable via setter
        engine.set_correlation_event_mode(CorrelationEventMode::Full);

        for i in 0..2 {
            let v = json!({"EventType": "login", "User": "admin"});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if i == 1 {
                assert_eq!(result.correlations.len(), 1);
                assert!(result.correlations[0].events.is_some());
                assert_eq!(result.correlations[0].events.as_ref().unwrap().len(), 2);
            }
        }
    }

    #[test]
    fn test_correlation_events_eviction_syncs_with_window() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 10s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Full,
            max_correlation_events: 100,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        // Push 2 events at ts=1000,1001 — within the 10s window
        for i in 0..2 {
            let v = json!({"EventType": "login", "User": "admin", "idx": i});
            let event = Event::from_value(&v);
            engine.process_event_at(&event, 1000 + i);
        }

        // Push 1 more event at ts=1015 — the first 2 events are now outside the
        // 10s window (cutoff = 1015 - 10 = 1005)
        let v = json!({"EventType": "login", "User": "admin", "idx": 2});
        let event = Event::from_value(&v);
        let result = engine.process_event_at(&event, 1015);
        // Should NOT fire: only 1 event in window (the one at ts=1015)
        assert!(
            result.correlations.is_empty(),
            "should not fire — old events evicted"
        );

        // Push 2 more to reach threshold
        for i in 3..5 {
            let v = json!({"EventType": "login", "User": "admin", "idx": i});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1016 + i - 3);
            if i == 4 {
                assert_eq!(result.correlations.len(), 1);
                let events = result.correlations[0].events.as_ref().unwrap();
                // Should have events from ts=1015,1016,1017 — not the old ones
                assert_eq!(events.len(), 3);
                for ev in events {
                    assert!(ev["idx"].as_i64().unwrap() >= 2);
                }
            }
        }
    }

    #[test]
    fn test_event_buffer_monitoring() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 100
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Full,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        assert_eq!(engine.event_buffer_count(), 0);
        assert_eq!(engine.event_buffer_bytes(), 0);

        // Push some events
        for i in 0..5 {
            let v = json!({"EventType": "login", "User": "admin"});
            let event = Event::from_value(&v);
            engine.process_event_at(&event, 1000 + i);
        }

        assert_eq!(engine.event_buffer_count(), 1); // one group key
        assert!(engine.event_buffer_bytes() > 0);
    }

    #[test]
    fn test_correlation_refs_mode_basic() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Refs,
            max_correlation_events: 10,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let mut corr_result = None;
        for i in 0..3 {
            let v = json!({"EventType": "login", "User": "admin", "id": format!("evt-{i}"), "@timestamp": 1000 + i});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if !result.correlations.is_empty() {
                corr_result = Some(result.correlations[0].clone());
            }
        }

        let result = corr_result.expect("correlation should have fired");
        // In refs mode: events should be None, event_refs should be Some
        assert!(
            result.events.is_none(),
            "Full events should not be included in refs mode"
        );
        let refs = result
            .event_refs
            .expect("event_refs should be present in refs mode");
        assert_eq!(refs.len(), 3);
        assert_eq!(refs[0].timestamp, 1000);
        assert_eq!(refs[0].id, Some("evt-0".to_string()));
        assert_eq!(refs[1].id, Some("evt-1".to_string()));
        assert_eq!(refs[2].id, Some("evt-2".to_string()));
    }

    #[test]
    fn test_correlation_refs_mode_no_id_field() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let config = CorrelationConfig {
            correlation_event_mode: CorrelationEventMode::Refs,
            ..Default::default()
        };
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let mut corr_result = None;
        for i in 0..2 {
            let v = json!({"EventType": "login", "User": "admin"});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if !result.correlations.is_empty() {
                corr_result = Some(result.correlations[0].clone());
            }
        }

        let result = corr_result.expect("correlation should have fired");
        let refs = result.event_refs.expect("event_refs should be present");
        // No ID field in events → id should be None
        for r in &refs {
            assert_eq!(r.id, None);
        }
    }

    #[test]
    fn test_per_correlation_custom_attributes_from_yaml() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
custom_attributes:
    rsigma.correlation_event_mode: refs
    rsigma.max_correlation_events: "5"
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 3
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        // Engine mode is None (default), but per-correlation should override to Refs
        let config = CorrelationConfig::default();
        let mut engine = CorrelationEngine::new(config);
        engine.add_collection(&collection).unwrap();

        let mut corr_result = None;
        for i in 0..3 {
            let v = json!({"EventType": "login", "User": "admin", "id": format!("e{i}")});
            let event = Event::from_value(&v);
            let result = engine.process_event_at(&event, 1000 + i);
            if !result.correlations.is_empty() {
                corr_result = Some(result.correlations[0].clone());
            }
        }

        let result = corr_result.expect("correlation should fire with per-correlation refs mode");
        // Per-correlation override should enable refs mode even though engine default is None
        assert!(result.events.is_none());
        let refs = result
            .event_refs
            .expect("event_refs via per-correlation override");
        assert_eq!(refs.len(), 3);
        assert_eq!(refs[0].id, Some("e0".to_string()));
    }

    #[test]
    fn test_per_correlation_custom_attr_suppress_and_action() {
        let yaml = r#"
title: Login
id: login-rule
logsource:
    category: auth
detection:
    selection:
        EventType: login
    condition: selection
---
title: Many Logins
custom_attributes:
    rsigma.suppress: 10s
    rsigma.action: reset
correlation:
    type: event_count
    rules:
        - login-rule
    group-by:
        - User
    timespan: 60s
    condition:
        gte: 2
level: high
"#;
        let collection = parse_sigma_yaml(yaml).unwrap();
        let mut engine = CorrelationEngine::new(CorrelationConfig::default());
        engine.add_collection(&collection).unwrap();

        // Verify the compiled correlation has per-rule overrides
        assert_eq!(engine.correlations[0].suppress_secs, Some(10));
        assert_eq!(
            engine.correlations[0].action,
            Some(CorrelationAction::Reset)
        );
    }
}