lean_ctx/core/

cognition_loop.rs

//! Hebbian-inspired Cognition Loop — periodic background reorganization of knowledge.
//! Runs 9 steps: seed promote, structural repair, fidelity check, lateral synthesis,
//! contradiction resolution, hebbian strengthen, decay, compact, observation synthesis.

use std::collections::HashSet;

use chrono::{Duration, Utc};

use crate::core::knowledge::ProjectKnowledge;
use crate::core::knowledge_relations::{
    KnowledgeEdgeKind, KnowledgeNodeRef, KnowledgeRelationGraph,
};
use crate::core::memory_policy::MemoryPolicy;

const LATERAL_SIM_THRESHOLD: f64 = 0.3;
const LATERAL_MAX_NEW_EDGES: usize = 20;
const HEBBIAN_CO_RETRIEVAL_HOURS: i64 = 1;
const EDGE_STALE_DAYS: i64 = 30;

// Observation synthesis (#802/cognition): how many facts an entity needs before it
// earns a summary, and the digest size caps that keep the value byte-stable.
const SYNTHESIS_MAX_MEMBERS: usize = 6;
const SYNTHESIS_VALUE_MAX: usize = 400;

#[derive(Debug, Clone, Default)]
pub struct CognitionLoopReport {
    pub steps_run: u8,
    pub facts_promoted: u32,
    pub edges_repaired: u32,
    pub edges_strengthened: u32,
    pub facts_decayed: u32,
    pub facts_archived: u32,
    pub contradictions_resolved: u32,
    pub lateral_connections: u32,
    /// Facts whose confidence was lifted by the replay-consolidation pass (#3).
    pub facts_consolidated: u32,
    /// Per-entity observation summaries written/refreshed by synthesis (#802).
    pub observations_synthesized: u32,
    pub duration_ms: u64,
}

impl std::fmt::Display for CognitionLoopReport {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(
            f,
            "Cognition Loop ({} steps, {}ms): promoted={}, repaired={}, \
             strengthened={}, decayed={}, archived={}, contradictions={}, lateral={}, \
             consolidated={}, synthesized={}",
            self.steps_run,
            self.duration_ms,
            self.facts_promoted,
            self.edges_repaired,
            self.edges_strengthened,
            self.facts_decayed,
            self.facts_archived,
            self.contradictions_resolved,
            self.lateral_connections,
            self.facts_consolidated,
            self.observations_synthesized,
        )
    }
}

pub fn run_cognition_loop(project_root: &str, max_steps: u8) -> CognitionLoopReport {
    let start = std::time::Instant::now();
    let mut report = CognitionLoopReport::default();

    let config = crate::core::config::Config::load();
    let Ok(policy) = config.memory_policy_effective() else {
        return report;
    };
    let synth_min_cluster = config.autonomy.cognition_synthesis_min_cluster.max(1);

    // Knowledge read-modify-write under the shared in-process + cross-process
    // lock so this loop (also driven by the background cognition scheduler)
    // never clobbers a concurrent foreground `remember`/`relate` write (issue
    // #326). The relation graph is loaded and saved inside the same critical
    // section; no step re-enters the knowledge lock, so this cannot deadlock.
    let _ = ProjectKnowledge::mutate_locked(project_root, |knowledge| {
        let project_hash = knowledge.project_hash.clone();
        let mut graph = KnowledgeRelationGraph::load_or_create(&project_hash);

        if max_steps >= 1 {
            report.facts_promoted = step_seed_promote(project_root, knowledge, &policy);
            report.steps_run = 1;
        }

        if max_steps >= 2 {
            report.edges_repaired = step_structural_repair(&mut graph, knowledge);
            report.steps_run = 2;
        }

        // Step 3: Fidelity Check (structural only, no LLM)
        if max_steps >= 3 {
            report.steps_run = 3;
        }

        if max_steps >= 4 {
            report.lateral_connections = step_lateral_synthesis(knowledge, &mut graph);
            report.steps_run = 4;
        }

        if max_steps >= 5 {
            report.contradictions_resolved = step_contradiction_resolution(knowledge);
            report.steps_run = 5;
        }

        if max_steps >= 6 {
            report.edges_strengthened = step_hebbian_strengthen(knowledge, &mut graph);
            report.steps_run = 6;
        }

        if max_steps >= 7 {
            report.facts_decayed = step_decay(knowledge, &mut graph, &policy);
            report.steps_run = 7;
        }

        if max_steps >= 8 {
            let lifecycle = knowledge.run_memory_lifecycle(&policy);
            report.facts_archived = lifecycle.archived_count as u32;
            // Step 8b (#3): complementary-learning-systems consolidation lifts the
            // confidence of related, frequently-retrieved facts.
            report.facts_consolidated = step_replay_consolidation(knowledge);
            if report.facts_consolidated > 0 {
                crate::core::introspect::tick("memory_consolidation");
            }
            report.steps_run = 8;
        }

        // Step 9 (#802): synthesize per-entity observation summaries from the now
        // settled store (after lifecycle), so summaries reflect surviving facts.
        if max_steps >= 9 {
            report.observations_synthesized =
                step_synthesize_observations(knowledge, &policy, synth_min_cluster);
            report.steps_run = 9;
        }

        let _ = graph.save();
    });

    report.duration_ms = start.elapsed().as_millis() as u64;
    report
}

/// Step 1: Promote recent session decisions/findings into project knowledge.
fn step_seed_promote(
    _project_root: &str,
    knowledge: &mut ProjectKnowledge,
    policy: &MemoryPolicy,
) -> u32 {
    let Some(session) = crate::core::session::SessionState::load_latest() else {
        return 0;
    };

    let mut count = 0u32;
    let max_decisions = 5usize;
    let max_findings = 8usize;

    let mut decisions = session.decisions.clone();
    decisions.sort_by_key(|d| std::cmp::Reverse(d.timestamp));
    decisions.truncate(max_decisions);
    for d in &decisions {
        let key = slug_key(&d.summary, 50);
        knowledge.remember("decision", &key, &d.summary, &session.id, 0.9, policy);
        count += 1;
    }

    let mut findings = session.findings.clone();
    findings.sort_by_key(|f| std::cmp::Reverse(f.timestamp));
    let mut kept = 0usize;
    for f in &findings {
        if kept >= max_findings {
            break;
        }
        if finding_salience(&f.summary) < 45 {
            continue;
        }
        let key = if let Some(ref file) = f.file {
            if let Some(line) = f.line {
                format!("{file}:{line}")
            } else {
                file.clone()
            }
        } else {
            format!("finding-{}", slug_key(&f.summary, 36))
        };
        knowledge.remember("finding", &key, &f.summary, &session.id, 0.75, policy);
        count += 1;
        kept += 1;
    }

    count
}

/// Step 2: Remove edges whose endpoints no longer exist in the knowledge store.
fn step_structural_repair(graph: &mut KnowledgeRelationGraph, knowledge: &ProjectKnowledge) -> u32 {
    let fact_ids: HashSet<String> = knowledge
        .facts
        .iter()
        .filter(|f| f.is_current())
        .map(|f| format!("{}/{}", f.category, f.key))
        .collect();

    let before = graph.edges.len();
    graph
        .edges
        .retain(|e| fact_ids.contains(&e.from.id()) && fact_ids.contains(&e.to.id()));
    (before - graph.edges.len()) as u32
}

/// Step 4: Connect related facts that share vocabulary but lack an explicit edge.
fn step_lateral_synthesis(knowledge: &ProjectKnowledge, graph: &mut KnowledgeRelationGraph) -> u32 {
    let current: Vec<_> = knowledge.facts.iter().filter(|f| f.is_current()).collect();

    let existing_pairs: HashSet<(String, String)> = graph
        .edges
        .iter()
        .map(|e| (e.from.id(), e.to.id()))
        .collect();

    let mut added = 0u32;

    for (i, a) in current.iter().enumerate() {
        if added >= LATERAL_MAX_NEW_EDGES as u32 {
            break;
        }
        for b in &current[i + 1..] {
            if added >= LATERAL_MAX_NEW_EDGES as u32 {
                break;
            }
            let id_a = format!("{}/{}", a.category, a.key);
            let id_b = format!("{}/{}", b.category, b.key);
            if existing_pairs.contains(&(id_a.clone(), id_b.clone()))
                || existing_pairs.contains(&(id_b.clone(), id_a.clone()))
            {
                continue;
            }
            let sim = crate::core::memory_consolidation::token_jaccard(&a.value, &b.value);
            if sim >= LATERAL_SIM_THRESHOLD {
                let from = KnowledgeNodeRef::new(&a.category, &a.key);
                let to = KnowledgeNodeRef::new(&b.category, &b.key);
                graph.upsert_edge(from, to, KnowledgeEdgeKind::RelatedTo, "cognition-loop");
                added += 1;
            }
        }
    }

    added
}

/// Step 5: Resolve contradictions — same category+key, different values.
/// Keeps the fact with higher quality_score, archives the other.
fn step_contradiction_resolution(knowledge: &mut ProjectKnowledge) -> u32 {
    let now = Utc::now();
    let mut resolved = 0u32;

    let mut seen: std::collections::HashMap<(String, String), usize> =
        std::collections::HashMap::new();
    let mut to_archive: Vec<usize> = Vec::new();

    for (i, f) in knowledge.facts.iter().enumerate() {
        if !f.is_current() {
            continue;
        }
        let key = (f.category.clone(), f.key.clone());
        if let Some(&prev_idx) = seen.get(&key) {
            let prev = &knowledge.facts[prev_idx];
            if prev.value != f.value {
                if prev.quality_score() >= f.quality_score() {
                    to_archive.push(i);
                } else {
                    to_archive.push(prev_idx);
                    seen.insert(key, i);
                }
                resolved += 1;
            }
        } else {
            seen.insert(key, i);
        }
    }

    for &idx in &to_archive {
        knowledge.facts[idx].valid_until = Some(now);
    }

    resolved
}

/// Step 6: Strengthen edges between facts co-retrieved in the same session window.
fn step_hebbian_strengthen(
    knowledge: &ProjectKnowledge,
    graph: &mut KnowledgeRelationGraph,
) -> u32 {
    let retrieved: Vec<_> = knowledge
        .facts
        .iter()
        .filter(|f| f.is_current() && f.last_retrieved.is_some())
        .collect();

    let window = Duration::hours(HEBBIAN_CO_RETRIEVAL_HOURS);
    let mut strengthened = 0u32;

    for (i, a) in retrieved.iter().enumerate() {
        let Some(a_time) = a.last_retrieved else {
            continue;
        };
        for b in &retrieved[i + 1..] {
            let Some(b_time) = b.last_retrieved else {
                continue;
            };
            let diff = (a_time - b_time).abs();
            if diff <= window {
                let from = KnowledgeNodeRef::new(&a.category, &a.key);
                let to = KnowledgeNodeRef::new(&b.category, &b.key);
                if !graph.strengthen_edge(&from, &to, 0.15) {
                    graph.upsert_edge(from, to, KnowledgeEdgeKind::RelatedTo, "hebbian");
                }
                strengthened += 1;
            }
        }
    }

    strengthened
}

/// Step 7: Decay confidence on stale facts, and decay edge counts for unseen edges.
fn step_decay(
    knowledge: &mut ProjectKnowledge,
    graph: &mut KnowledgeRelationGraph,
    policy: &MemoryPolicy,
) -> u32 {
    let lifecycle_cfg = crate::core::memory_lifecycle::LifecycleConfig {
        max_facts: policy.knowledge.max_facts,
        decay_rate_per_day: policy.lifecycle.decay_rate,
        low_confidence_threshold: policy.lifecycle.low_confidence_threshold,
        stale_days: policy.lifecycle.stale_days,
        consolidation_similarity: policy.lifecycle.similarity_threshold,
        forgetting_model: crate::core::memory_lifecycle::ForgettingModel::parse(
            &policy.lifecycle.forgetting_model,
        ),
        base_stability_days: policy.lifecycle.base_stability_days,
        archetype_aware_decay: policy.lifecycle.archetype_aware_decay,
    };
    crate::core::memory_lifecycle::apply_confidence_decay(&mut knowledge.facts, &lifecycle_cfg);

    let low_conf_count = knowledge
        .facts
        .iter()
        .filter(|f| f.is_current() && f.confidence < 0.3)
        .count() as u32;

    graph.decay_all_edges(1.0);
    graph.prune_weak_edges(0.05);

    let stale_cutoff = Utc::now() - Duration::days(EDGE_STALE_DAYS);
    graph.edges.retain_mut(|e| {
        let last = e.last_seen.unwrap_or(e.created_at);
        if last < stale_cutoff {
            if e.count <= 1 {
                return false;
            }
            e.count = e.count.saturating_sub(1);
        }
        true
    });

    low_conf_count
}

/// Step 8b (#3): replay consolidation over the knowledge facts. Maps facts into
/// consolidation entries, runs the sleep-inspired NREM/REM/replay pass, then
/// promotes the replay-boosted importance back onto fact confidence. Additive:
/// merges and pruning are owned by the lifecycle step, so here we only *lift*
/// the confidence of facts the replay pass found related-and-co-accessed —
/// never lower or delete. Deterministic.
fn step_replay_consolidation(knowledge: &mut ProjectKnowledge) -> u32 {
    use crate::core::memory_consolidation::{KnowledgeEntry, consolidate};

    let mut entries: Vec<KnowledgeEntry> = knowledge
        .facts
        .iter()
        .filter(|f| f.is_current())
        .map(|f| {
            let last_access = f
                .last_retrieved
                .unwrap_or(f.last_confirmed)
                .timestamp()
                .max(0) as u64;
            KnowledgeEntry {
                key: format!("{}/{}", f.category, f.key),
                content: f.value.clone(),
                access_count: u64::from(f.retrieval_count),
                last_access,
                created_at: f.created_at.timestamp().max(0) as u64,
                importance: f64::from(f.confidence),
            }
        })
        .collect();
    if entries.len() < 2 {
        return 0;
    }
    consolidate(&mut entries);

    let boosted: std::collections::HashMap<String, f64> =
        entries.into_iter().map(|e| (e.key, e.importance)).collect();

    let mut promoted = 0u32;
    for f in knowledge.facts.iter_mut().filter(|f| f.is_current()) {
        let id = format!("{}/{}", f.category, f.key);
        if let Some(&imp) = boosted.get(&id) {
            let new_conf = (imp as f32).min(1.0);
            if new_conf > f.confidence + 0.001 {
                f.confidence = new_conf;
                promoted += 1;
            }
        }
    }
    promoted
}

/// Idle replay (#7): the sleep-inspired (sharp-wave-ripple) consolidation pass,
/// run when the agent has been quiet rather than as part of the periodic loop.
/// Reloads knowledge under the shared lock, replays the consolidation/promote
/// step, and reports the facts whose confidence the replay lifted. Distinct from
/// the in-loop step (#3) so idle-time "rest" consolidation is observable on its
/// own via `introspect`. Deterministic; mutates the store, never tool output.
pub fn run_idle_replay(project_root: &str) -> u32 {
    let mut promoted = 0u32;
    let _ = ProjectKnowledge::mutate_locked(project_root, |knowledge| {
        promoted = step_replay_consolidation(knowledge);
    });
    if promoted > 0 {
        crate::core::introspect::tick("replay_consolidation");
    }
    promoted
}

/// Step 9 (#802/cognition): deterministically synthesize per-entity *observation*
/// summaries from clusters of related raw facts — lean-ctx's take on Hindsight's
/// observation network. Current facts (never synthesized observations) are grouped
/// by an entity anchor (a file path referenced in the key/value, else the
/// category); each cluster of `>= min_cluster` facts writes/refreshes one compact
/// observation via [`ProjectKnowledge::remember`] (idempotent + versioned), so
/// summaries never summarize summaries.
///
/// Deterministic: the value is a stable function of the source facts' content
/// (sorted, capped, char-boundary-truncated — no timestamps/counters), so it never
/// perturbs the prompt cache (#498). Runs in the background loop, never a hot path.
fn step_synthesize_observations(
    knowledge: &mut ProjectKnowledge,
    policy: &MemoryPolicy,
    min_cluster: usize,
) -> u32 {
    use std::collections::BTreeMap;

    if min_cluster == 0 {
        return 0;
    }

    // Cluster current facts by entity. Only *synthesized* observations are skipped
    // (no recursion) — raw user findings (also `Observation` archetype) are valid
    // input. BTreeMap → deterministic entity order.
    let mut clusters: BTreeMap<String, Vec<(String, String, f32)>> = BTreeMap::new();
    for f in knowledge.facts.iter().filter(|f| f.is_current()) {
        if f.is_synthesized_observation() {
            continue;
        }
        let entity = synthesis_entity_anchor(&f.category, &f.key, &f.value);
        clusters.entry(entity).or_default().push((
            f.category.clone(),
            f.value.clone(),
            f.confidence,
        ));
    }

    let mut count = 0u32;
    for (entity, mut members) in clusters {
        if members.len() < min_cluster {
            continue;
        }
        // Strongest evidence first, then lexical — deterministic.
        members.sort_by(|a, b| {
            b.2.partial_cmp(&a.2)
                .unwrap_or(std::cmp::Ordering::Equal)
                .then_with(|| a.1.cmp(&b.1))
        });
        members.truncate(SYNTHESIS_MAX_MEMBERS);

        let summary = synthesize_observation_value(&entity, &members);
        // Optional LLM refinement (opt-in via `llm.enabled`); deterministic fallback.
        let summary = crate::core::llm_enhance::enhance_observation(&entity, &summary);
        // An observation never out-confidences its own evidence: mean, capped.
        let mean = members.iter().map(|m| m.2).sum::<f32>() / members.len() as f32;
        knowledge.remember(
            "observation",
            &entity,
            &summary,
            crate::core::knowledge::COGNITION_SYNTHESIS_SOURCE,
            mean.min(0.9),
            policy,
        );
        count += 1;
    }

    if count > 0 {
        crate::core::introspect::tick("observation_synthesis");
    }
    count
}

/// The entity an observation summarizes: the first file path referenced in the
/// fact key (findings key by `file:line`) or value, else the category. Deterministic.
fn synthesis_entity_anchor(category: &str, key: &str, value: &str) -> String {
    crate::core::content_chunk::extract_file_references(key)
        .into_iter()
        .next()
        .or_else(|| {
            crate::core::content_chunk::extract_file_references(value)
                .into_iter()
                .next()
        })
        .unwrap_or_else(|| category.to_string())
}

/// Compose a deterministic, structured digest of an entity's facts grouped by their
/// source category. Char-boundary-truncated so the stored value is byte-stable.
fn synthesize_observation_value(entity: &str, members: &[(String, String, f32)]) -> String {
    use std::collections::BTreeMap;
    let mut by_cat: BTreeMap<&str, Vec<&str>> = BTreeMap::new();
    for (cat, val, _) in members {
        by_cat.entry(cat.as_str()).or_default().push(val.as_str());
    }
    let body = by_cat
        .into_iter()
        .map(|(cat, vals)| format!("{cat}: {}", vals.join("; ")))
        .collect::<Vec<_>>()
        .join(" | ");
    format!(
        "{entity} — {}",
        truncate_on_char_boundary(&body, SYNTHESIS_VALUE_MAX)
    )
}

/// Truncate to at most `max` bytes on a UTF-8 boundary, appending an ellipsis when
/// it actually shortens. Deterministic; used to bound synthesized observation text.
fn truncate_on_char_boundary(s: &str, max: usize) -> String {
    if s.len() <= max {
        return s.to_string();
    }
    let mut end = max;
    while end > 0 && !s.is_char_boundary(end) {
        end -= 1;
    }
    format!("{}…", &s[..end])
}

fn slug_key(s: &str, max: usize) -> String {
    let mut out = String::new();
    for ch in s.chars() {
        if out.len() >= max {
            break;
        }
        if ch.is_ascii_alphanumeric() {
            out.push(ch.to_ascii_lowercase());
        } else if (ch.is_whitespace() || ch == '-' || ch == '_')
            && !out.ends_with('-')
            && !out.is_empty()
        {
            out.push('-');
        }
    }
    out.trim_matches('-').to_string()
}

fn finding_salience(summary: &str) -> u32 {
    let s = summary.to_lowercase();
    let mut score = 20u32;
    let boosts = [
        ("error", 25),
        ("failed", 25),
        ("panic", 30),
        ("assert", 20),
        ("forbidden", 25),
        ("timeout", 20),
        ("deadlock", 25),
        ("security", 25),
        ("vuln", 25),
        ("e0", 15),
    ];
    for (pat, b) in boosts {
        if s.contains(pat) {
            score = score.saturating_add(b);
        }
    }
    score
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::core::knowledge::KnowledgeArchetype;
    use crate::core::knowledge_relations::KnowledgeEdge;
    use crate::core::memory_boundary::FactPrivacy;

    fn make_fact(
        category: &str,
        key: &str,
        value: &str,
        confidence: f32,
    ) -> crate::core::knowledge::KnowledgeFact {
        crate::core::knowledge::KnowledgeFact {
            category: category.to_string(),
            key: key.to_string(),
            value: value.to_string(),
            source_session: "test".to_string(),
            confidence,
            created_at: Utc::now(),
            last_confirmed: Utc::now(),
            retrieval_count: 0,
            last_retrieved: None,
            valid_from: Some(Utc::now()),
            valid_until: None,
            supersedes: None,
            confirmation_count: 1,
            feedback_up: 0,
            feedback_down: 0,
            last_feedback: None,
            privacy: FactPrivacy::default(),
            sensitivity: crate::core::sensitivity::SensitivityLevel::default(),
            imported_from: None,
            archetype: KnowledgeArchetype::default(),
            fidelity: None,
            revision_count: 0,
        }
    }

    fn make_retrieved_fact(
        category: &str,
        key: &str,
        value: &str,
        retrieved_at: chrono::DateTime<Utc>,
    ) -> crate::core::knowledge::KnowledgeFact {
        let mut f = make_fact(category, key, value, 0.9);
        f.last_retrieved = Some(retrieved_at);
        f.retrieval_count = 1;
        f
    }

    fn make_knowledge(
        project_root: &str,
        facts: Vec<crate::core::knowledge::KnowledgeFact>,
    ) -> ProjectKnowledge {
        ProjectKnowledge {
            project_root: project_root.to_string(),
            project_hash: "test-hash".to_string(),
            facts,
            patterns: Vec::new(),
            history: Vec::new(),
            updated_at: Utc::now(),
            judged_pairs: Vec::new(),
        }
    }

    fn make_graph(edges: Vec<KnowledgeEdge>) -> KnowledgeRelationGraph {
        KnowledgeRelationGraph {
            project_hash: "test-hash".to_string(),
            edges,
            updated_at: Utc::now(),
        }
    }

    fn make_edge(from_cat: &str, from_key: &str, to_cat: &str, to_key: &str) -> KnowledgeEdge {
        KnowledgeEdge {
            from: KnowledgeNodeRef::new(from_cat, from_key),
            to: KnowledgeNodeRef::new(to_cat, to_key),
            kind: KnowledgeEdgeKind::RelatedTo,
            created_at: Utc::now(),
            last_seen: Some(Utc::now()),
            count: 1,
            source_session: "test".to_string(),
            strength: 0.5,
            decay_rate: 0.02,
        }
    }

    #[test]
    fn structural_repair_removes_orphaned_edges() {
        let knowledge = make_knowledge(
            "/tmp/test",
            vec![
                make_fact("arch", "db", "PostgreSQL", 0.9),
                make_fact("arch", "cache", "Redis", 0.8),
            ],
        );

        let mut graph = make_graph(vec![
            make_edge("arch", "db", "arch", "cache"),
            make_edge("arch", "db", "arch", "nonexistent"),
            make_edge("gone", "missing", "arch", "db"),
        ]);

        let removed = step_structural_repair(&mut graph, &knowledge);
        assert_eq!(removed, 2);
        assert_eq!(graph.edges.len(), 1);
        assert_eq!(graph.edges[0].from.key, "db");
        assert_eq!(graph.edges[0].to.key, "cache");
    }

    #[test]
    fn lateral_synthesis_connects_similar_facts() {
        let knowledge = make_knowledge(
            "/tmp/test",
            vec![
                make_fact(
                    "arch",
                    "db",
                    "PostgreSQL database primary storage backend",
                    0.9,
                ),
                make_fact("arch", "cache", "Redis cache for sessions", 0.8),
                make_fact(
                    "deploy",
                    "db-host",
                    "PostgreSQL database primary storage on AWS",
                    0.7,
                ),
            ],
        );

        let mut graph = make_graph(Vec::new());
        let added = step_lateral_synthesis(&knowledge, &mut graph);

        assert!(
            added >= 1,
            "Should connect facts sharing vocabulary (PostgreSQL database primary storage)"
        );
        assert!(
            graph.edges.iter().any(|e| {
                (e.from.key == "db" && e.to.key == "db-host")
                    || (e.from.key == "db-host" && e.to.key == "db")
            }),
            "Should have edge between db and db-host"
        );
    }

    #[test]
    fn contradiction_resolution_keeps_higher_quality() {
        let mut f1 = make_fact("arch", "db", "PostgreSQL", 0.9);
        f1.confirmation_count = 3;
        let f2 = make_fact("arch", "db", "MySQL", 0.5);

        let mut knowledge = make_knowledge("/tmp/test", vec![f1, f2]);
        let resolved = step_contradiction_resolution(&mut knowledge);

        assert_eq!(resolved, 1);
        let current: Vec<_> = knowledge.facts.iter().filter(|f| f.is_current()).collect();
        assert_eq!(current.len(), 1);
        assert_eq!(current[0].value, "PostgreSQL");
    }

    #[test]
    fn hebbian_strengthen_co_retrieval() {
        let now = Utc::now();
        let knowledge = make_knowledge(
            "/tmp/test",
            vec![
                make_retrieved_fact("arch", "db", "PostgreSQL", now),
                make_retrieved_fact("arch", "cache", "Redis", now - Duration::minutes(30)),
                make_retrieved_fact("arch", "queue", "Kafka", now - Duration::hours(5)),
            ],
        );

        let mut graph = make_graph(Vec::new());
        let strengthened = step_hebbian_strengthen(&knowledge, &mut graph);

        assert!(
            strengthened >= 1,
            "Should strengthen co-retrieved facts within 1h window"
        );
        let has_db_cache = graph.edges.iter().any(|e| {
            (e.from.key == "db" && e.to.key == "cache")
                || (e.from.key == "cache" && e.to.key == "db")
        });
        assert!(has_db_cache, "db and cache were retrieved within 1h");
    }

    #[test]
    fn decay_reduces_stale_edge_counts() {
        let old = Utc::now() - Duration::days(45);
        let mut graph = make_graph(vec![
            {
                let mut e = make_edge("arch", "db", "arch", "cache");
                e.last_seen = Some(old);
                e.count = 3;
                e
            },
            {
                let mut e = make_edge("arch", "old", "arch", "ancient");
                e.last_seen = Some(old);
                e.count = 1;
                e
            },
        ]);

        let policy = MemoryPolicy::default();
        let mut knowledge = make_knowledge(
            "/tmp/test",
            vec![
                make_fact("arch", "db", "PostgreSQL", 0.9),
                make_fact("arch", "cache", "Redis", 0.8),
            ],
        );

        step_decay(&mut knowledge, &mut graph, &policy);

        assert_eq!(
            graph.edges.len(),
            1,
            "Edge with count=1 and stale should be removed"
        );
        assert_eq!(
            graph.edges[0].count, 2,
            "Edge with count=3 should be decremented to 2"
        );
    }

    #[test]
    fn replay_consolidation_promotes_related_accessed_facts() {
        // #3: related (jaccard in replay band) + frequently-retrieved facts get
        // their confidence lifted by the replay-boost pass.
        let mut f1 = make_fact(
            "arch",
            "db",
            "uses postgres database for primary storage",
            0.5,
        );
        f1.retrieval_count = 50;
        f1.last_retrieved = Some(Utc::now());
        let mut f2 = make_fact(
            "arch",
            "db2",
            "uses postgres database for sessions cache",
            0.5,
        );
        f2.retrieval_count = 50;
        f2.last_retrieved = Some(Utc::now());

        let mut knowledge = make_knowledge("/tmp/test", vec![f1, f2]);
        let promoted = step_replay_consolidation(&mut knowledge);
        assert!(
            promoted >= 1,
            "related, frequently-accessed facts should be promoted (#3)"
        );
        assert!(
            knowledge.facts.iter().any(|f| f.confidence > 0.5),
            "confidence must be lifted by replay boost"
        );
    }

    #[test]
    fn idle_replay_consolidates_from_disk() {
        // #7: the idle replay pass loads knowledge under lock, consolidates the
        // related/frequently-retrieved facts, and persists the lifted confidence.
        let _lock = crate::core::data_dir::test_env_lock();
        let tmp = tempfile::tempdir().expect("tempdir");
        crate::test_env::set_var(
            "LEAN_CTX_DATA_DIR",
            tmp.path().to_string_lossy().to_string(),
        );
        let project_root = tmp.path().join("proj");
        std::fs::create_dir_all(&project_root).expect("mkdir");
        let root = project_root.to_string_lossy().to_string();

        let policy = MemoryPolicy::default();
        let mut knowledge = ProjectKnowledge::load_or_create(&root);
        knowledge.remember(
            "arch",
            "db",
            "uses postgres database for primary storage",
            "s1",
            0.5,
            &policy,
        );
        knowledge.remember(
            "arch",
            "db2",
            "uses postgres database for sessions cache",
            "s1",
            0.5,
            &policy,
        );
        for f in &mut knowledge.facts {
            f.retrieval_count = 50;
            f.last_retrieved = Some(Utc::now());
        }
        let _ = knowledge.save();

        let promoted = run_idle_replay(&root);
        assert!(
            promoted >= 1,
            "idle replay should consolidate related facts (#7)"
        );

        crate::test_env::remove_var("LEAN_CTX_DATA_DIR");
    }

    #[test]
    fn cognition_loop_runs_all_steps() {
        let _lock = crate::core::data_dir::test_env_lock();
        let tmp = tempfile::tempdir().expect("tempdir");
        crate::test_env::set_var(
            "LEAN_CTX_DATA_DIR",
            tmp.path().to_string_lossy().to_string(),
        );

        let project_root = tmp.path().join("proj");
        std::fs::create_dir_all(&project_root).expect("mkdir");
        let project_root_str = project_root.to_string_lossy().to_string();

        let policy = MemoryPolicy::default();
        let mut knowledge = ProjectKnowledge::load_or_create(&project_root_str);
        knowledge.remember("arch", "db", "PostgreSQL", "s1", 0.9, &policy);
        knowledge.remember("arch", "cache", "Redis", "s1", 0.8, &policy);
        knowledge.remember("deploy", "host", "AWS", "s1", 0.7, &policy);
        let _ = knowledge.save();

        let report = run_cognition_loop(&project_root_str, 8);
        assert_eq!(report.steps_run, 8);

        crate::test_env::remove_var("LEAN_CTX_DATA_DIR");
    }

    #[test]
    fn synthesize_observation_value_is_deterministic() {
        let members = vec![
            ("finding".to_string(), "b issue".to_string(), 0.5f32),
            ("finding".to_string(), "a issue".to_string(), 0.9f32),
            ("gotcha".to_string(), "race".to_string(), 0.7f32),
        ];
        let v1 = synthesize_observation_value("src/x.rs", &members);
        let v2 = synthesize_observation_value("src/x.rs", &members);
        assert_eq!(v1, v2, "synthesis value must be deterministic");
        assert!(v1.starts_with("src/x.rs — "));
        // Grouped by source category in deterministic (BTreeMap) order.
        let f = v1.find("finding:").expect("finding group");
        let g = v1.find("gotcha:").expect("gotcha group");
        assert!(f < g, "categories grouped in sorted order");
    }

    #[test]
    fn synthesis_entity_anchor_resolves_file_then_category() {
        assert_eq!(
            synthesis_entity_anchor("finding", "src/auth.rs:42", "x"),
            "src/auth.rs"
        );
        assert_eq!(
            synthesis_entity_anchor("decision", "no-file", "see src/lib.rs here"),
            "src/lib.rs"
        );
        assert_eq!(
            synthesis_entity_anchor("decision", "plain-key", "no path at all"),
            "decision"
        );
    }

    #[test]
    fn step_synthesizes_per_entity_observation_and_is_idempotent() {
        let _lock = crate::core::data_dir::test_env_lock();
        let tmp = tempfile::tempdir().expect("tempdir");
        crate::test_env::set_var(
            "LEAN_CTX_DATA_DIR",
            tmp.path().to_string_lossy().to_string(),
        );

        let policy = MemoryPolicy::default();
        let mut k = ProjectKnowledge::new("/tmp/test-synthesis");
        // Three facts anchored to the same file → one entity cluster.
        k.remember(
            "finding",
            "src/auth.rs:10",
            "missing null check",
            "s1",
            0.8,
            &policy,
        );
        k.remember(
            "finding",
            "src/auth.rs:20",
            "token not validated",
            "s1",
            0.7,
            &policy,
        );
        k.remember(
            "gotcha",
            "src/auth.rs:30",
            "race on refresh",
            "s1",
            0.9,
            &policy,
        );

        let made = step_synthesize_observations(&mut k, &policy, 3);
        assert_eq!(made, 1, "one observation for the clustered entity");

        let obs: Vec<_> = k
            .facts
            .iter()
            .filter(|f| f.is_current() && f.is_synthesized_observation())
            .collect();
        assert_eq!(obs.len(), 1);
        assert_eq!(obs[0].key, "src/auth.rs");
        assert_eq!(obs[0].archetype, KnowledgeArchetype::Observation);

        // Re-run with unchanged facts → same value → confirmation, not a duplicate.
        let again = step_synthesize_observations(&mut k, &policy, 3);
        assert_eq!(again, 1, "step still writes (confirms) the summary");
        let current = k
            .facts
            .iter()
            .filter(|f| f.is_current() && f.is_synthesized_observation())
            .count();
        assert_eq!(current, 1, "idempotent: no duplicate observation");

        crate::test_env::remove_var("LEAN_CTX_DATA_DIR");
    }

    #[test]
    fn synthesis_excludes_synthesized_observations_from_input() {
        let policy = MemoryPolicy::default();
        let mut k = ProjectKnowledge::new("/tmp/test-no-recursion");
        // A pre-existing synthesized observation must never be re-summarized.
        k.remember(
            "observation",
            "src/x.rs",
            "src/x.rs — finding: a; b",
            crate::core::knowledge::COGNITION_SYNTHESIS_SOURCE,
            0.6,
            &policy,
        );
        k.remember("finding", "src/y.rs:1", "issue one", "s1", 0.5, &policy);
        // Only one non-synthesized entity ("src/y.rs"), below the threshold → none.
        let made = step_synthesize_observations(&mut k, &policy, 3);
        assert_eq!(made, 0, "no entity reaches the cluster threshold");
    }

    #[test]
    fn cognition_loop_step_9_synthesizes_observations() {
        let _lock = crate::core::data_dir::test_env_lock();
        let tmp = tempfile::tempdir().expect("tempdir");
        crate::test_env::set_var(
            "LEAN_CTX_DATA_DIR",
            tmp.path().to_string_lossy().to_string(),
        );
        let project_root = tmp.path().join("proj");
        std::fs::create_dir_all(&project_root).expect("mkdir");
        let root = project_root.to_string_lossy().to_string();

        let policy = MemoryPolicy::default();
        let mut knowledge = ProjectKnowledge::load_or_create(&root);
        knowledge.remember(
            "finding",
            "src/api.rs:1",
            "no auth on route",
            "s1",
            0.8,
            &policy,
        );
        knowledge.remember(
            "finding",
            "src/api.rs:2",
            "missing rate limit",
            "s1",
            0.7,
            &policy,
        );
        knowledge.remember(
            "gotcha",
            "src/api.rs:3",
            "panics on empty body",
            "s1",
            0.9,
            &policy,
        );
        let _ = knowledge.save();

        let report = run_cognition_loop(&root, 9);
        assert_eq!(report.steps_run, 9);
        assert!(
            report.observations_synthesized >= 1,
            "step 9 must synthesize at least one observation"
        );

        let reloaded = ProjectKnowledge::load_or_create(&root);
        assert!(
            reloaded
                .facts
                .iter()
                .any(|f| f.is_current() && f.is_synthesized_observation()),
            "synthesized observation must persist"
        );

        crate::test_env::remove_var("LEAN_CTX_DATA_DIR");
    }
}