mentedb_cognitive/

write_inference.rs

use mentedb_core::MemoryNode;
use mentedb_core::edge::EdgeType;
use mentedb_core::memory::MemoryType;
use mentedb_core::types::MemoryId;

use crate::llm::{
    CognitiveLlmService, ContradictionVerdict, InvalidationVerdict, LlmJudge, MemorySummary,
};

#[derive(Debug, Clone)]
pub enum InferredAction {
    FlagContradiction {
        existing: MemoryId,
        new: MemoryId,
        reason: String,
    },
    MarkObsolete {
        memory: MemoryId,
        superseded_by: MemoryId,
    },
    /// Set valid_until on the old memory instead of deleting it.
    InvalidateMemory {
        memory: MemoryId,
        superseded_by: MemoryId,
        valid_until: u64,
    },
    /// A byte-identical re-save: invalidate the duplicate and link it to the
    /// surviving copy with a `Derived` edge (deduplication lineage), NOT a
    /// `Supersedes` edge. A Supersedes edge here reads as a memory superseding an
    /// exact copy of itself, which only shows up as noise in the conflict view.
    DeduplicateExact {
        duplicate: MemoryId,
        keeper: MemoryId,
    },
    /// Update the content of an existing memory with merged information.
    UpdateContent {
        memory: MemoryId,
        new_content: String,
        reason: String,
    },
    CreateEdge {
        source: MemoryId,
        target: MemoryId,
        edge_type: EdgeType,
        weight: f32,
    },
    UpdateConfidence {
        memory: MemoryId,
        new_confidence: f32,
    },
    PropagateBeliefChange {
        root: MemoryId,
        delta: f32,
    },
}

fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
    if a.len() != b.len() || a.is_empty() {
        return 0.0;
    }
    let mut dot = 0.0f32;
    let mut norm_a = 0.0f32;
    let mut norm_b = 0.0f32;
    for i in 0..a.len() {
        dot += a[i] * b[i];
        norm_a += a[i] * a[i];
        norm_b += b[i] * b[i];
    }
    let denom = norm_a.sqrt() * norm_b.sqrt();
    if denom == 0.0 { 0.0 } else { dot / denom }
}

/// Configuration for write-time inference thresholds.
#[derive(Debug, Clone)]
pub struct WriteInferenceConfig {
    /// Similarity above which two memories may contradict (default: 0.95).
    pub contradiction_threshold: f32,
    /// Similarity above which an older memory is marked obsolete (default: 0.85).
    pub obsolete_threshold: f32,
    /// Minimum similarity for creating a Related edge (default: 0.6).
    pub related_min: f32,
    /// Maximum similarity for creating a Related edge (default: 0.85).
    pub related_max: f32,
    /// Minimum similarity for a Correction to supersede (default: 0.5).
    pub correction_threshold: f32,
    /// Multiplier applied to original confidence on correction (default: 0.5).
    pub confidence_decay_factor: f32,
    /// Minimum confidence after decay (default: 0.1).
    pub confidence_floor: f32,
    /// Whether the value-update rule runs: a new memory that shares an older
    /// same-owner memory's sentence frame (long common token prefix) but ends
    /// in a different value supersedes it, so "favorite coffee is a cortado"
    /// followed by "favorite coffee is a flat white" keeps only the correction.
    /// This is the cheap, LLM-free slice of contradiction handling; unlike bare
    /// cosine (which cannot tell paraphrase from contradiction and was removed
    /// at ~0% precision), the shared-frame test only fires on the update shape.
    pub value_update_enabled: bool,
    /// Minimum embedding similarity for the value-update rule (default: 0.3).
    ///
    /// This is a loose topical floor, not the primary gate. A value correction
    /// inherently drives cosine down (the changed value is most of a short
    /// fact's meaning), so real corrections land well below a high cosine bar:
    /// measured on Titan V2 at 256 dims, "favorite coffee is a cortado" vs
    /// "... a flat white" is 0.61 and "phone is 1234" vs "... 4321" is 0.51;
    /// on OpenAI text-embedding-3-small they are 0.80 and 0.79. A 0.88 gate
    /// (the first default) was unreachable on either embedder, and even 0.5
    /// sat inside the real-pair range and silently gated identical-frame
    /// pairs in production (observed on Cohere 1024: "uses Next.js for the
    /// front end" vs "uses React for the front end" never superseded). The
    /// embedder-independent `is_value_update` shared frame test carries the
    /// precision here; the floor only rejects genuinely off-topic pairs that
    /// a shared token frame collided on, which sit near 0.2 and below.
    pub value_update_min_similarity: f32,
    /// Minimum fraction of the longer memory's tokens that must be shared
    /// frame, the common prefix plus the common suffix around the changed
    /// value (default: 0.6).
    pub value_update_prefix_share: f32,
    /// Maximum differing value tokens per side (default: 4). Larger middles
    /// mean genuinely different statements, not a value change.
    pub value_update_max_tail: usize,
}

impl Default for WriteInferenceConfig {
    fn default() -> Self {
        Self {
            contradiction_threshold: 0.95,
            obsolete_threshold: 0.85,
            related_min: 0.6,
            related_max: 0.85,
            correction_threshold: 0.5,
            confidence_decay_factor: 0.5,
            confidence_floor: 0.1,
            value_update_enabled: true,
            value_update_min_similarity: 0.3,
            value_update_prefix_share: 0.6,
            value_update_max_tail: 4,
        }
    }
}

/// Lowercased alphanumeric tokens of a memory's content, the unit of the
/// value-update frame comparison.
fn frame_tokens(s: &str) -> Vec<String> {
    s.split(|c: char| !c.is_alphanumeric())
        .filter(|t| !t.is_empty())
        .map(|t| t.to_lowercase())
        .collect()
}

/// True when `new` and `old` share the same sentence frame with a short
/// differing value somewhere inside it: a common token prefix and a common
/// token suffix that together cover most of the sentence, with a small
/// non-identical middle on each side. The value may sit at the end ("the
/// user's favorite coffee order is a cortado" vs "... is a flat white") or
/// mid sentence ("i use next.js for the front end" vs "i use react for the
/// front end", where the value is wrapped by a shared tail). A subject swap
/// at the head ("my dog is allergic to chicken" vs "my cat is allergic to
/// chicken") shares at most one leading token, so the two token minimum
/// prefix keeps it out: changing the subject is a different fact, not a
/// value update. The original prefix-only test could not see shared tails,
/// which made every mid sentence value change invisible and left
/// contradicting facts coexisting.
///
/// Mid sentence middles that are purely numeric on both sides ("rule 3" vs
/// "rule 4", "room 2" vs "room 5") are identifiers naming different things,
/// not changed values, and such facts legitimately coexist; only the end
/// position keeps numeric updates ("phone is 1234" vs "phone is 4321").
/// Precision beats recall here on purpose: a wrong supersession silently
/// deletes a true fact, while a missed one is caught by the LLM
/// contradiction layer.
fn is_value_update(new: &str, old: &str, frame_share: f32, max_value: usize) -> bool {
    let nt = frame_tokens(new);
    let ot = frame_tokens(old);
    if nt.is_empty() || ot.is_empty() || nt == ot {
        return false;
    }
    // Very short sentences cannot carry enough frame to tell an update from
    // an enumeration: "directive number 1" and "directive number 2" are two
    // items, not one changed value. Real value sentences name a relation and
    // a value and run longer ("my phone number is 4321" is five tokens).
    if nt.len().min(ot.len()) < 4 {
        return false;
    }
    let prefix = nt.iter().zip(ot.iter()).take_while(|(a, b)| a == b).count();
    if prefix < 2 {
        return false;
    }
    let suffix = nt
        .iter()
        .rev()
        .zip(ot.iter().rev())
        .take_while(|(a, b)| a == b)
        .count()
        // The suffix may not reach back into the prefix on the shorter side.
        .min(nt.len().min(ot.len()) - prefix);
    let longest = nt.len().max(ot.len());
    let value_new = nt.len() - prefix - suffix;
    let value_old = ot.len() - prefix - suffix;
    if suffix > 0 {
        let all_numeric =
            |tokens: &[String]| tokens.iter().all(|t| t.chars().all(|c| c.is_ascii_digit()));
        if all_numeric(&nt[prefix..nt.len() - suffix])
            && all_numeric(&ot[prefix..ot.len() - suffix])
        {
            return false;
        }
    }
    (prefix + suffix) as f32 / longest as f32 >= frame_share
        && value_new.max(value_old) > 0
        && value_new <= max_value
        && value_old <= max_value
}

/// Decide which of two value-frame-matching memories supersedes the other, and
/// which is left stale. Order-independent: a `Correction` supersedes a plain
/// fact whichever was stored first (the extractor types a correction explicitly,
/// so it is the authoritative value); otherwise the later-created memory wins.
/// Returns `(winner_id, loser_id, loser_confidence)`, or `None` when there is no
/// correction signal and the timestamps are equal, in which case the two are
/// kept as coexisting facts rather than guessing a direction.
fn value_update_resolution(
    new: &MemoryNode,
    existing: &MemoryNode,
) -> Option<(MemoryId, MemoryId, f32)> {
    let new_correction = new.memory_type == MemoryType::Correction;
    let old_correction = existing.memory_type == MemoryType::Correction;
    if new_correction && !old_correction {
        return Some((new.id, existing.id, existing.confidence));
    }
    if old_correction && !new_correction {
        return Some((existing.id, new.id, new.confidence));
    }
    if new.created_at > existing.created_at {
        return Some((new.id, existing.id, existing.confidence));
    }
    if existing.created_at > new.created_at {
        return Some((existing.id, new.id, new.confidence));
    }
    None
}

pub struct WriteInferenceEngine {
    config: WriteInferenceConfig,
}

impl WriteInferenceEngine {
    pub fn new() -> Self {
        Self {
            config: WriteInferenceConfig::default(),
        }
    }

    pub fn with_config(config: WriteInferenceConfig) -> Self {
        Self { config }
    }

    /// Whether `new` and `old` look like a value update (same sentence frame,
    /// changed value tail) under this engine's configured thresholds. A
    /// deduplication pass that merges near-duplicates uses this as a guard so a
    /// correction ("coffee is a cortado" then "coffee is a flat white") is
    /// never collapsed into the fact it corrects; the two must stay distinct
    /// for supersession to resolve them.
    pub fn looks_like_value_update(&self, new: &str, old: &str) -> bool {
        is_value_update(
            new,
            old,
            self.config.value_update_prefix_share,
            self.config.value_update_max_tail,
        )
    }

    pub fn infer_on_write(
        &self,
        new_memory: &MemoryNode,
        existing_memories: &[MemoryNode],
        existing_edges: &[(MemoryId, MemoryId, EdgeType)],
    ) -> Vec<InferredAction> {
        let _ = existing_edges; // reserved for future graph-aware inference
        let mut actions = Vec::new();
        // Value-frame-matched pairs are owned by the value-update rule below; the
        // broad correction handler skips them so a pair is never resolved twice.
        let mut value_update_handled: Vec<MemoryId> = Vec::new();

        for existing in existing_memories {
            if existing.id == new_memory.id {
                continue;
            }

            let sim = cosine_similarity(&new_memory.embedding, &existing.embedding);

            // Cosine similarity alone cannot tell a reworded duplicate from a
            // genuine contradiction: paraphrases ("uses Postgres" / "uses
            // PostgreSQL") and opposites ("prefers tabs" / "prefers spaces")
            // both land in the high-similarity band. So the cheap write-time
            // heuristic makes only the two judgments similarity CAN support:
            //   1. byte-identical content -> a true duplicate, supersede it
            //   2. moderate similarity    -> a Related edge
            // Every semantic decision (merge, paraphrase dedup, real
            // contradiction, which-fact-wins supersession) is deferred to the
            // LLM paths (`consolidate_memories`, `detect_conflicts_with_llm`),
            // which read the actual text. Flagging contradictions from bare
            // similarity here produced ~0% precision (every reworded re-save
            // looked like a contradiction), so it was removed.
            if existing.content == new_memory.content
                && sim > self.config.obsolete_threshold
                && new_memory.created_at > existing.created_at
            {
                actions.push(InferredAction::DeduplicateExact {
                    duplicate: existing.id,
                    keeper: new_memory.id,
                });
            } else if self.config.value_update_enabled
                && existing.agent_id == new_memory.agent_id
                && existing.user_id == new_memory.user_id
                && is_value_update(
                    &new_memory.content,
                    &existing.content,
                    self.config.value_update_prefix_share,
                    self.config.value_update_max_tail,
                )
            {
                if sim < self.config.value_update_min_similarity {
                    // A frame matched pair below the floor is exactly the
                    // signal that the floor sits wrong for the embedding
                    // space in use; log it so production data keeps the
                    // default honest.
                    tracing::info!(
                        similarity = sim,
                        "value update frame matched below the cosine floor, skipped"
                    );
                    continue;
                }
                // Same sentence frame, different value: one supersedes the other.
                // The value-update rule owns this pair (so the correction handler
                // skips it), and the winner is decided deterministically rather
                // than by arrival order: a correction beats a plain fact, else the
                // later-created wins. Batch enrichment stores extracted facts in
                // arbitrary order with near-equal timestamps, so an order-dependent
                // rule fired only about half the time; this resolves the same way
                // whether the correction arrives before or after the original.
                value_update_handled.push(existing.id);
                if let Some((winner, loser, loser_confidence)) =
                    value_update_resolution(new_memory, existing)
                {
                    // Invalidate the loser, not just edge it: recall's
                    // validity filter reads valid_until, so a Supersedes edge
                    // alone left the stale value recallable and the agent
                    // kept answering with it. InvalidateMemory sets
                    // valid_until and creates the Supersedes edge.
                    let now = std::time::SystemTime::now()
                        .duration_since(std::time::UNIX_EPOCH)
                        .unwrap_or_default()
                        .as_micros() as u64;
                    actions.push(InferredAction::InvalidateMemory {
                        memory: loser,
                        superseded_by: winner,
                        valid_until: now,
                    });
                    actions.push(InferredAction::UpdateConfidence {
                        memory: loser,
                        new_confidence: (loser_confidence * self.config.confidence_decay_factor)
                            .max(self.config.confidence_floor),
                    });
                }
                // else: no correction signal and equal timestamps, so the two are
                // kept as coexisting facts (a multi-valued attribute) rather than
                // one silently superseding the other.
            } else if sim > self.config.related_min && sim <= self.config.related_max {
                actions.push(InferredAction::CreateEdge {
                    source: new_memory.id,
                    target: existing.id,
                    edge_type: EdgeType::Related,
                    weight: sim,
                });
            }
        }

        // Correction type: find the most similar existing memory and supersede
        // it, skipping any pair the value-update rule already resolved above.
        if new_memory.memory_type == MemoryType::Correction
            && let Some(original) = existing_memories
                .iter()
                .filter(|m| m.id != new_memory.id && !value_update_handled.contains(&m.id))
                .max_by(|a, b| {
                    cosine_similarity(&new_memory.embedding, &a.embedding)
                        .partial_cmp(&cosine_similarity(&new_memory.embedding, &b.embedding))
                        .unwrap_or(std::cmp::Ordering::Equal)
                })
        {
            let sim = cosine_similarity(&new_memory.embedding, &original.embedding);
            if sim > self.config.correction_threshold {
                actions.push(InferredAction::CreateEdge {
                    source: new_memory.id,
                    target: original.id,
                    edge_type: EdgeType::Supersedes,
                    weight: 1.0,
                });
                actions.push(InferredAction::UpdateConfidence {
                    memory: original.id,
                    new_confidence: (original.confidence * self.config.confidence_decay_factor)
                        .max(self.config.confidence_floor),
                });
            }
        }

        actions
    }

    /// LLM enhanced write inference. Uses the CognitiveLlmService when available
    /// to make smarter invalidation and contradiction decisions. Falls back to
    /// cosine similarity heuristics for moderate similarity (Related edges) and
    /// correction handling.
    pub async fn infer_on_write_with_llm<J: LlmJudge>(
        &self,
        new_memory: &MemoryNode,
        existing_memories: &[MemoryNode],
        existing_edges: &[(MemoryId, MemoryId, EdgeType)],
        llm: &CognitiveLlmService<J>,
    ) -> Vec<InferredAction> {
        let _ = existing_edges;
        let mut actions = Vec::new();

        let new_summary = memory_to_summary(new_memory);

        for existing in existing_memories {
            if existing.id == new_memory.id {
                continue;
            }

            let sim = cosine_similarity(&new_memory.embedding, &existing.embedding);

            // Only consult LLM for memories with moderate+ similarity (> 0.5)
            // to avoid burning tokens on completely unrelated pairs
            if sim > 0.5 && existing.agent_id == new_memory.agent_id {
                let old_summary = memory_to_summary(existing);

                // LLM invalidation check
                if let Ok(verdict) = llm.judge_invalidation(&old_summary, &new_summary).await {
                    match verdict {
                        InvalidationVerdict::Invalidate { reason: _ } => {
                            // InvalidateMemory sets valid_until and creates the
                            // Supersedes edge; no separate MarkObsolete/CreateEdge.
                            actions.push(InferredAction::InvalidateMemory {
                                memory: existing.id,
                                superseded_by: new_memory.id,
                                valid_until: new_memory.created_at,
                            });
                            actions.push(InferredAction::UpdateConfidence {
                                memory: existing.id,
                                new_confidence: (existing.confidence
                                    * self.config.confidence_decay_factor)
                                    .max(self.config.confidence_floor),
                            });
                            // Skip contradiction check since we already know the relationship
                            continue;
                        }
                        InvalidationVerdict::Update {
                            merged_content,
                            reason,
                        } => {
                            actions.push(InferredAction::UpdateContent {
                                memory: existing.id,
                                new_content: merged_content,
                                reason,
                            });
                            continue;
                        }
                        InvalidationVerdict::Keep { .. } => {
                            // Fall through to contradiction check
                        }
                    }
                }

                // LLM contradiction check for high similarity pairs that weren't invalidated
                if sim > 0.7
                    && existing.content != new_memory.content
                    && let Ok(verdict) = llm.detect_contradiction(&old_summary, &new_summary).await
                {
                    match verdict {
                        ContradictionVerdict::Contradicts { reason } => {
                            actions.push(InferredAction::FlagContradiction {
                                existing: existing.id,
                                new: new_memory.id,
                                reason,
                            });
                        }
                        ContradictionVerdict::Supersedes { winner, reason: _ } => {
                            let winner_is_new = winner == new_memory.id.to_string();
                            let (obsolete, superseder) = if winner_is_new {
                                (existing.id, new_memory.id)
                            } else {
                                (new_memory.id, existing.id)
                            };
                            actions.push(InferredAction::InvalidateMemory {
                                memory: obsolete,
                                superseded_by: superseder,
                                valid_until: new_memory.created_at,
                            });
                        }
                        ContradictionVerdict::Compatible { .. } => {}
                    }
                }
            }

            // Moderate similarity: create Related edge (heuristic, no LLM needed)
            if sim > self.config.related_min && sim <= self.config.related_max {
                actions.push(InferredAction::CreateEdge {
                    source: new_memory.id,
                    target: existing.id,
                    edge_type: EdgeType::Related,
                    weight: sim,
                });
            }
        }

        // Correction type still uses heuristic (find most similar and supersede)
        if new_memory.memory_type == MemoryType::Correction
            && let Some(original) = existing_memories
                .iter()
                .filter(|m| m.id != new_memory.id)
                .max_by(|a, b| {
                    cosine_similarity(&new_memory.embedding, &a.embedding)
                        .partial_cmp(&cosine_similarity(&new_memory.embedding, &b.embedding))
                        .unwrap_or(std::cmp::Ordering::Equal)
                })
        {
            let sim = cosine_similarity(&new_memory.embedding, &original.embedding);
            if sim > self.config.correction_threshold {
                actions.push(InferredAction::InvalidateMemory {
                    memory: original.id,
                    superseded_by: new_memory.id,
                    valid_until: new_memory.created_at,
                });
                actions.push(InferredAction::UpdateConfidence {
                    memory: original.id,
                    new_confidence: (original.confidence * self.config.confidence_decay_factor)
                        .max(self.config.confidence_floor),
                });
            }
        }

        actions
    }
}

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

fn memory_to_summary(m: &MemoryNode) -> MemorySummary {
    MemorySummary {
        id: m.id,
        content: m.content.clone(),
        memory_type: m.memory_type,
        confidence: m.confidence,
        created_at: m.created_at,
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use mentedb_core::memory::MemoryType;
    use mentedb_core::types::AgentId;

    use crate::llm::MockLlmJudge;

    fn make_memory(content: &str, embedding: Vec<f32>, mem_type: MemoryType) -> MemoryNode {
        let mut m = MemoryNode::new(AgentId::new(), mem_type, content.to_string(), embedding);
        m.created_at = 1000;
        m
    }

    #[test]
    fn test_heuristic_does_not_flag_contradiction_from_similarity() {
        // Two facts with high embedding similarity but different content look
        // identical to cosine similarity whether they are paraphrases or
        // opposites. The cheap heuristic must NOT guess "contradiction" here
        // (that produced ~0% precision in production); real contradiction
        // detection is the LLM path's job. High-sim different-content yields no
        // action from the heuristic.
        let agent = AgentId::new();
        let mut existing =
            make_memory("uses PostgreSQL", vec![1.0, 0.0, 0.0], MemoryType::Semantic);
        existing.agent_id = agent;

        let mut new_mem = make_memory("uses MySQL", vec![0.99, 0.01, 0.0], MemoryType::Semantic);
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            !actions
                .iter()
                .any(|a| matches!(a, InferredAction::FlagContradiction { .. })),
            "Heuristic must not flag contradictions from bare similarity, got: {:?}",
            actions
        );
    }

    #[test]
    fn value_update_supersedes_the_old_fact() {
        // The canonical correction shape: same sentence frame, new value at the
        // tail. The newer memory must supersede the older one so recall serves
        // only the corrected fact (live bug: both coffee facts coexisted).
        let agent = AgentId::new();
        let mut existing = make_memory(
            "The user's favorite coffee order is a cortado",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;
        let mut new_mem = make_memory(
            "The user's favorite coffee order is a flat white",
            vec![0.98, 0.02, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing.clone()], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::InvalidateMemory { memory, superseded_by, .. }
                    if *superseded_by == new_mem.id && *memory == existing.id
            )),
            "value update must invalidate the stale fact, got: {actions:?}"
        );
        assert!(
            actions
                .iter()
                .any(|a| matches!(a, InferredAction::UpdateConfidence { memory, .. } if *memory == existing.id)),
            "superseded fact's confidence must decay"
        );
    }

    #[test]
    fn value_update_fires_at_realistic_embedder_cosine() {
        // Regression for the production-dead rule: real embedders put a value
        // correction far below a high cosine bar (measured: Titan V2 0.61,
        // OpenAI 3-small 0.80 for cortado vs flat white; 0.51 / 0.79 for a phone
        // number change). The old 0.88 default was unreachable on either, so the
        // rule never fired in prod. A frame-matched correction at cosine 0.6 must
        // now supersede; the structural shared-frame test, not the cosine, is
        // what proves it is a correction.
        let agent = AgentId::new();
        let mut existing = make_memory(
            "The user's favorite coffee order is a cortado",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;
        // cosine([1,0,0], [0.6,0.8,0]) = 0.6: below the old 0.88 gate, above 0.5.
        let mut new_mem = make_memory(
            "The user's favorite coffee order is a flat white",
            vec![0.6, 0.8, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let actions =
            WriteInferenceEngine::new().infer_on_write(&new_mem, &[existing.clone()], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::InvalidateMemory { memory, superseded_by, .. }
                    if *superseded_by == new_mem.id && *memory == existing.id
            )),
            "a frame-matched correction at realistic cosine 0.6 must supersede, got: {actions:?}"
        );
    }

    #[test]
    fn value_update_fires_below_the_old_floor() {
        // The production miss on Cohere 1024: an identical frame pair whose
        // cosine sits in the 0.3 to 0.5 band was silently gated by the 0.5
        // floor. cosine([1,0,0],[0.4,0.9165,0]) = 0.4 must fire now.
        let agent = AgentId::new();
        let mut existing = make_memory(
            "The user uses Next.js for the front end",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;
        let mut new_mem = make_memory(
            "The user uses React for the front end",
            vec![0.4, 0.9165, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let actions =
            WriteInferenceEngine::new().infer_on_write(&new_mem, &[existing.clone()], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::InvalidateMemory { memory, superseded_by, .. }
                    if *superseded_by == new_mem.id && *memory == existing.id
            )),
            "an identical frame pair at cosine 0.4 must supersede, got: {actions:?}"
        );
    }

    #[test]
    fn value_update_below_cosine_floor_does_not_supersede() {
        // The cosine floor still guards: a shared token frame that collided on
        // genuinely off-topic content (cosine 0.2, below the 0.3 floor) must not
        // supersede, even though the frame test alone would pass. Structure is
        // primary, but the floor rejects pathological frame collisions.
        let agent = AgentId::new();
        let mut existing = make_memory(
            "The user's favorite coffee order is a cortado",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;
        // cosine([1,0,0], [0.2,0.9798,0]) = 0.2: below the 0.3 floor.
        let mut new_mem = make_memory(
            "The user's favorite coffee order is a flat white",
            vec![0.2, 0.9798, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let actions = WriteInferenceEngine::new().infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            !actions
                .iter()
                .any(|a| matches!(a, InferredAction::InvalidateMemory { .. })),
            "below the cosine floor a frame match must not supersede, got: {actions:?}"
        );
    }

    #[test]
    fn correction_supersedes_regardless_of_store_order() {
        // Batch enrichment can store the correction BEFORE the original it
        // corrects. The original arriving after the existing correction must
        // still be the one marked stale: the correction wins either way. This is
        // the case the old order-dependent rule missed (both facts coexisted).
        let agent = AgentId::new();
        let mut correction = make_memory(
            "The user's favorite coffee is a flat white",
            vec![1.0, 0.0, 0.0],
            MemoryType::Correction,
        );
        correction.agent_id = agent;
        correction.created_at = 5000;
        let mut original = make_memory(
            "The user's favorite coffee is a cortado",
            vec![0.98, 0.02, 0.0],
            MemoryType::Semantic,
        );
        original.agent_id = agent;
        original.created_at = 5000; // same batch timestamp

        // The original arrives as the new memory; the correction already exists.
        let actions =
            WriteInferenceEngine::new().infer_on_write(&original, &[correction.clone()], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::InvalidateMemory { memory, superseded_by, .. }
                    if *superseded_by == correction.id && *memory == original.id
            )),
            "the correction must invalidate the later-arriving original, got: {actions:?}"
        );
    }

    #[test]
    fn correction_supersedes_original_at_equal_timestamps() {
        // The other store order: original first, correction second, same batch
        // timestamp. The correction still wins.
        let agent = AgentId::new();
        let mut original = make_memory(
            "The user's favorite coffee is a cortado",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        original.agent_id = agent;
        original.created_at = 5000;
        let mut correction = make_memory(
            "The user's favorite coffee is a flat white",
            vec![0.98, 0.02, 0.0],
            MemoryType::Correction,
        );
        correction.agent_id = agent;
        correction.created_at = 5000;

        let actions =
            WriteInferenceEngine::new().infer_on_write(&correction, &[original.clone()], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::InvalidateMemory { memory, superseded_by, .. }
                    if *superseded_by == correction.id && *memory == original.id
            )),
            "the correction must invalidate the original at equal timestamps, got: {actions:?}"
        );
    }

    #[test]
    fn multivalued_facts_at_equal_timestamps_are_both_kept() {
        // Two plain facts sharing a frame that are not a correction of each other
        // (a multi-valued attribute) must not supersede when timestamps are equal,
        // so making the rule order-independent does not start eating coexisting
        // facts. Only a correction signal or a clear time order triggers it.
        let agent = AgentId::new();
        let mut a = make_memory(
            "The user knows Rust",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        a.agent_id = agent;
        a.created_at = 5000;
        let mut b = make_memory(
            "The user knows Python",
            vec![0.97, 0.03, 0.0],
            MemoryType::Semantic,
        );
        b.agent_id = agent;
        b.created_at = 5000;

        let actions = WriteInferenceEngine::new().infer_on_write(&b, &[a], &[]);
        assert!(
            !actions.iter().any(|x| matches!(
                x,
                InferredAction::CreateEdge {
                    edge_type: EdgeType::Supersedes,
                    ..
                }
            )),
            "coexisting facts at equal timestamps must not supersede, got: {actions:?}"
        );
    }

    #[test]
    fn subject_change_is_not_a_value_update() {
        // "my dog ..." vs "my cat ..." differ at the head: two facts about two
        // subjects, both must survive.
        let agent = AgentId::new();
        let mut existing = make_memory(
            "My dog is allergic to chicken",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;
        let mut new_mem = make_memory(
            "My cat is allergic to chicken",
            vec![0.99, 0.01, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            !actions.iter().any(|a| matches!(
                a,
                InferredAction::CreateEdge {
                    edge_type: EdgeType::Supersedes,
                    ..
                }
            )),
            "a subject change must not supersede, got: {actions:?}"
        );
    }

    #[test]
    fn value_update_respects_owner_and_config() {
        // Cross-owner pairs never supersede, and the rule can be disabled.
        let mut existing = make_memory(
            "The user's favorite coffee order is a cortado",
            vec![1.0, 0.0, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = AgentId::new();
        let mut new_mem = make_memory(
            "The user's favorite coffee order is a flat white",
            vec![0.98, 0.02, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = AgentId::new(); // different owner
        new_mem.created_at = 2000;

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing.clone()], &[]);
        assert!(
            !actions.iter().any(|a| matches!(
                a,
                InferredAction::CreateEdge {
                    edge_type: EdgeType::Supersedes,
                    ..
                }
            )),
            "cross-owner memories must never supersede each other"
        );

        // Same owner but rule disabled: no supersession either.
        let agent = AgentId::new();
        existing.agent_id = agent;
        new_mem.agent_id = agent;
        let engine_off = WriteInferenceEngine::with_config(WriteInferenceConfig {
            value_update_enabled: false,
            ..WriteInferenceConfig::default()
        });
        let actions = engine_off.infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            !actions.iter().any(|a| matches!(
                a,
                InferredAction::CreateEdge {
                    edge_type: EdgeType::Supersedes,
                    ..
                }
            )),
            "disabled rule must not supersede"
        );
    }

    #[test]
    fn is_value_update_frame_analysis() {
        // Positive: long shared prefix, short differing tail.
        assert!(is_value_update(
            "my phone number is 4321",
            "my phone number is 1234",
            0.6,
            4
        ));
        // Negative: head differs (different subject).
        assert!(!is_value_update(
            "my cat is allergic to chicken",
            "my dog is allergic to chicken",
            0.6,
            4
        ));
        // Negative: identical content is not an update.
        assert!(!is_value_update("same fact", "same fact", 0.6, 4));
        // Positive: the value sits mid sentence, wrapped by a shared tail.
        // This is the reported production miss: "i use next.js" then "now i
        // use react" coexisted forever because the old prefix-only test
        // could not see the shared "for the front end" tail.
        assert!(is_value_update(
            "The user uses React for the front end",
            "The user uses Next.js for the front end",
            0.6,
            4
        ));
        assert!(is_value_update(
            "i use react for the front end",
            "i use next.js for the front end",
            0.6,
            4
        ));
        // Negative: subject swap with a long shared tail must still not fire;
        // only a one token head is shared.
        assert!(!is_value_update(
            "the cat is allergic to chicken and needs the special food",
            "the dog is allergic to chicken and needs the special food",
            0.6,
            4
        ));
        // Negative: purely numeric mid sentence middles are identifiers
        // naming different things (rule 3 and rule 4 coexist), never a
        // changed value.
        assert!(!is_value_update(
            "API sibling rule 1 about payload naming",
            "API sibling rule 0 about payload naming",
            0.6,
            4
        ));
        // Positive is unaffected at the end position: a numeric tail is a
        // value ("phone is 1234" to "phone is 4321"), covered above.
        // Negative: three token enumerations are items, never updates.
        assert!(!is_value_update(
            "Directive number 1",
            "Directive number 0",
            0.6,
            4
        ));
        // Negative: tail too long (a genuinely different statement).
        assert!(!is_value_update(
            "the deploy runs every friday and sarah reviews it after standup",
            "the deploy runs every friday unless the release train is frozen for the quarter end audit",
            0.6,
            4
        ));
    }

    #[test]
    fn test_byte_identical_duplicate_is_deduplicated() {
        // Identical text is deduplicated (invalidate + Derived lineage), not a
        // supersession, so it never surfaces as a conflict.
        let agent = AgentId::new();
        let mut existing = make_memory(
            "Ran command: ls -la",
            vec![1.0, 0.0, 0.0],
            MemoryType::Episodic,
        );
        existing.agent_id = agent;

        let mut new_mem = make_memory(
            "Ran command: ls -la",
            vec![1.0, 0.0, 0.0],
            MemoryType::Episodic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            actions
                .iter()
                .any(|a| matches!(a, InferredAction::DeduplicateExact { .. })),
            "Expected identical duplicate to be deduplicated, got: {:?}",
            actions
        );
    }

    #[test]
    fn test_moderate_similarity_creates_edge() {
        let existing = make_memory("topic A", vec![1.0, 0.0, 0.0], MemoryType::Semantic);
        // ~0.7 similarity
        let new_mem = make_memory("topic B", vec![0.7, 0.714, 0.0], MemoryType::Semantic);

        let engine = WriteInferenceEngine::new();
        let actions = engine.infer_on_write(&new_mem, &[existing], &[]);
        assert!(
            actions.iter().any(|a| matches!(
                a,
                InferredAction::CreateEdge {
                    edge_type: EdgeType::Related,
                    ..
                }
            )),
            "Expected CreateEdge Related, got: {:?}",
            actions
        );
    }

    #[tokio::test]
    async fn test_llm_invalidation_emits_temporal_actions() {
        let agent = AgentId::new();
        let mut existing = make_memory(
            "Alice works at Acme",
            vec![0.8, 0.6, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;

        let mut new_mem = make_memory(
            "Alice joined Google last week",
            vec![0.75, 0.65, 0.1],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let judge =
            MockLlmJudge::new(r#"{"verdict": "invalidate", "reason": "Alice changed jobs"}"#);
        let llm = CognitiveLlmService::new(judge);
        let engine = WriteInferenceEngine::new();
        let actions = engine
            .infer_on_write_with_llm(&new_mem, &[existing.clone()], &[], &llm)
            .await;

        assert!(
            actions
                .iter()
                .any(|a| matches!(a, InferredAction::InvalidateMemory { .. })),
            "Expected InvalidateMemory from LLM verdict, got: {:?}",
            actions
        );
        // Exactly one invalidation action: InvalidateMemory both sets
        // valid_until and creates the Supersedes edge, so a MarkObsolete
        // alongside it would duplicate the edge.
        let invalidation_count = actions
            .iter()
            .filter(|a| {
                matches!(
                    a,
                    InferredAction::InvalidateMemory { .. } | InferredAction::MarkObsolete { .. }
                )
            })
            .count();
        assert_eq!(
            invalidation_count, 1,
            "Expected a single invalidation action, got: {:?}",
            actions
        );
    }

    #[tokio::test]
    async fn test_llm_update_emits_update_content() {
        let agent = AgentId::new();
        let mut existing = make_memory(
            "Project uses React",
            vec![0.8, 0.6, 0.0],
            MemoryType::Semantic,
        );
        existing.agent_id = agent;

        let mut new_mem = make_memory(
            "Project migrated from React to Vue",
            vec![0.75, 0.65, 0.1],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let judge = MockLlmJudge::new(
            r#"{"verdict": "update", "merged_content": "Project migrated from React to Vue in Q2", "reason": "adds temporal context"}"#,
        );
        let llm = CognitiveLlmService::new(judge);
        let engine = WriteInferenceEngine::new();
        let actions = engine
            .infer_on_write_with_llm(&new_mem, &[existing.clone()], &[], &llm)
            .await;

        assert!(
            actions
                .iter()
                .any(|a| matches!(a, InferredAction::UpdateContent { .. })),
            "Expected UpdateContent from LLM update verdict, got: {:?}",
            actions
        );
    }

    #[tokio::test]
    async fn test_llm_keep_falls_through_to_contradiction_check() {
        let agent = AgentId::new();
        // High similarity pair where LLM says keep but content differs
        let mut existing = make_memory("Prefers tabs", vec![0.9, 0.44, 0.0], MemoryType::Semantic);
        existing.agent_id = agent;

        let mut new_mem = make_memory(
            "Prefers spaces",
            vec![0.88, 0.47, 0.0],
            MemoryType::Semantic,
        );
        new_mem.agent_id = agent;
        new_mem.created_at = 2000;

        let judge = MockLlmJudge::new(
            r#"{"verdict": "compatible", "reason": "different formatting preferences"}"#,
        );
        let llm = CognitiveLlmService::new(judge);
        let engine = WriteInferenceEngine::new();
        let actions = engine
            .infer_on_write_with_llm(&new_mem, &[existing.clone()], &[], &llm)
            .await;

        // MockLlmJudge returns the same response for all calls, so the keep
        // from invalidation falls through, then contradiction returns compatible
        assert!(
            !actions
                .iter()
                .any(|a| matches!(a, InferredAction::FlagContradiction { .. })),
            "Should not flag contradiction when LLM says compatible",
        );
    }
}