src/mnemocore/core/engine.py

"""
Holographic Active Inference Memory Engine (HAIM) - Phase 4.3+
Uses Binary HDV for efficient storage and computation.
"""

from typing import List, Tuple, Dict, Optional, Any, TYPE_CHECKING, Deque

if TYPE_CHECKING:
    from .container import Container
    from .qdrant_store import QdrantStore
import heapq
from collections import deque
from itertools import islice
import numpy as np
import hashlib
import os
import json
import asyncio
import functools
import uuid
import re
from datetime import datetime, timezone
from loguru import logger

from .config import get_config, HAIMConfig, SubconsciousAIConfig
from .binary_hdv import BinaryHDV, TextEncoder, majority_bundle
from .node import MemoryNode
from .synapse import SynapticConnection
from .holographic import ConceptualMemory
from .tier_manager import TierManager

# Phase 4.0 imports
from .attention import XORAttentionMasker, AttentionConfig, XORIsolationMask, IsolationConfig
from .bayesian_ltp import get_bayesian_updater
from .semantic_consolidation import SemanticConsolidationWorker, SemanticConsolidationConfig
from .immunology import ImmunologyLoop, ImmunologyConfig
from .gap_detector import GapDetector, GapDetectorConfig
from .gap_filler import GapFiller, GapFillerConfig
from .synapse_index import SynapseIndex
from .subconscious_ai import SubconsciousAIWorker

# Phase 5 AGI Stores
from .working_memory import WorkingMemoryService
from .episodic_store import EpisodicStoreService
from .semantic_store import SemanticStoreService

# Phase 4.5: Recursive Synthesis Engine
from .recursive_synthesizer import RecursiveSynthesizer, SynthesizerConfig

# Observability imports (Phase 4.1)
from .metrics import (
    timer, traced, get_trace_id, set_trace_id,
    STORE_DURATION_SECONDS, QUERY_DURATION_SECONDS,
    MEMORY_COUNT_TOTAL, QUEUE_LENGTH, ERROR_TOTAL,
    update_memory_count, update_queue_length, record_error
)


class HAIMEngine:
    """
    Holographic Active Inference Memory Engine (Phase 4.3+)
    Uses Binary HDV and Tiered Storage for efficient cognitive memory.
    """

    @staticmethod
    @functools.lru_cache(maxsize=10000)
    def _get_token_vector(token: str, dimension: int) -> np.ndarray:
        """Cached generation of deterministic token vectors (legacy compatibility)."""
        seed_bytes = hashlib.shake_256(token.encode()).digest(4)
        seed = int.from_bytes(seed_bytes, 'little')
        return np.random.RandomState(seed).choice([-1, 1], size=dimension)

    def __init__(
        self,
        dimension: int = 16384,
        persist_path: Optional[str] = None,
        config: Optional[HAIMConfig] = None,
        tier_manager: Optional[TierManager] = None,
        working_memory: Optional[WorkingMemoryService] = None,
        episodic_store: Optional[EpisodicStoreService] = None,
        semantic_store: Optional[SemanticStoreService] = None,
    ):
        """
        Initialize HAIMEngine with optional dependency injection.

        Args:
            dimension: Vector dimensionality (default 16384).
            persist_path: Path to memory persistence file.
            config: Configuration object. If None, uses global get_config().
            tier_manager: TierManager instance. If None, creates a new one.
            working_memory: Optional Phase 5 WM service.
            episodic_store: Optional Phase 5 EM service.
            semantic_store: Optional Phase 5 Semantic service.
        """
        self.config = config or get_config()
        self.dimension = self.config.dimensionality

        # Initialization guard
        self._initialized: bool = False

        # Core Components
        self.tier_manager = tier_manager or TierManager(config=self.config)
        
        # Phase 5 Components
        self.working_memory = working_memory
        self.episodic_store = episodic_store
        self.semantic_store = semantic_store
        self.binary_encoder = TextEncoder(self.dimension)

        # ── Phase 3.x: synapse raw dicts (kept for backward compat) ──
        self.synapses: Dict[Tuple[str, str], SynapticConnection] = {}
        self.synapse_adjacency: Dict[str, List[SynapticConnection]] = {}
        # Async locks – safe to create here in Python 3.10+
        self.synapse_lock: asyncio.Lock = asyncio.Lock()
        # Serialises concurrent _save_synapses disk writes
        self._write_lock: asyncio.Lock = asyncio.Lock()
        # Serialises store-path persistence and episodic-chain updates
        self._store_lock: asyncio.Lock = asyncio.Lock()
        # Semaphore: only one dream cycle at a time (rate limiting)
        self._dream_sem: asyncio.Semaphore = asyncio.Semaphore(1)

        # ── Phase 4.0: hardened O(1) synapse adjacency index ──────────
        self._synapse_index = SynapseIndex()

        # ── Phase 4.0: XOR attention masker ───────────────────────────
        self.attention_masker = XORAttentionMasker(AttentionConfig())

        # ── Phase 4.1: XOR project isolation masker ───────────────────
        isolation_enabled = getattr(self.config, 'attention_masking', None)
        isolation_enabled = isolation_enabled.enabled if isolation_enabled else True
        self.isolation_masker = XORIsolationMask(IsolationConfig(
            enabled=isolation_enabled,
            dimension=self.dimension,
        ))

        # ── Phase 4.0: gap detector & filler (wired in initialize()) ──
        self.gap_detector = GapDetector(GapDetectorConfig())
        self._gap_filler: Optional[GapFiller] = None

        # ── Phase 4.0: semantic consolidation worker ───────────────────
        self._semantic_worker: Optional[SemanticConsolidationWorker] = None

        # ── Phase 4.0: immunology loop ─────────────────────────────────
        self._immunology: Optional[ImmunologyLoop] = None

        # ── Phase 4.4: subconscious AI worker (BETA) ───────────────────
        self._subconscious_ai: Optional[SubconsciousAIWorker] = None

        # ── Phase 4.5: recursive synthesizer ───────────────────────────
        self._recursive_synthesizer: Optional[RecursiveSynthesizer] = None
        
        # ── Phase 12.2: Contextual Topic Tracker ───────────────────────
        from .topic_tracker import TopicTracker
        self.topic_tracker = TopicTracker(self.config.context, self.dimension)

        # ── Phase 12.3: Preference Learning ────────────────────────────
        from .preference_store import PreferenceStore
        self.preference_store = PreferenceStore(self.config.preference, self.dimension)

        # ── Phase 13.2: Anticipatory Memory ────────────────────────────
        from .anticipatory import AnticipatoryEngine
        self.anticipatory_engine = AnticipatoryEngine(
            self.config.anticipatory, 
            self._synapse_index, 
            self.tier_manager, 
            self.topic_tracker
        )

        # Conceptual Layer (VSA Soul)
        data_dir = self.config.paths.data_dir
        self.soul = ConceptualMemory(dimension=self.dimension, storage_dir=data_dir)

        # Persistence paths
        self.persist_path = persist_path or self.config.paths.memory_file
        self.synapse_path = self.config.paths.synapses_file

        # Passive Subconscious Layer (bounded if configured)
        queue_maxlen = self.config.dream_loop.subconscious_queue_maxlen
        self.subconscious_queue: Deque[str] = deque(maxlen=queue_maxlen)
        self._last_stored_id: Optional[str] = None

        # Epistemic Drive
        self.epistemic_drive_active = True
        self.surprise_threshold = 0.7

    async def initialize(self):
        """Async initialization."""
        if self._initialized:
            return

        await self.tier_manager.initialize()
        await self._load_legacy_if_needed()
        await self._load_synapses()
        self._initialized = True

        # ── Phase 4.0: start background workers ───────────────────────
        self._semantic_worker = SemanticConsolidationWorker(self)
        await self._semantic_worker.start()

        self._immunology = ImmunologyLoop(self)
        await self._immunology.start()

        # ── Phase 4.4: start subconscious AI worker (if enabled) ──────
        if self.config.subconscious_ai.enabled:
            self._subconscious_ai = SubconsciousAIWorker(self, self.config.subconscious_ai)
            await self._subconscious_ai.start()
            logger.info("Phase 4.4 SubconsciousAI worker started (BETA).")

        logger.info("Phase 4.0 background workers started (consolidation + immunology).")

    async def _run_in_thread(self, func, *args, **kwargs):
        """Run blocking function in thread pool."""
        loop = asyncio.get_running_loop()
        return await loop.run_in_executor(None, functools.partial(func, *args, **kwargs))

    def calculate_eig(self, candidate: BinaryHDV, context: BinaryHDV) -> float:
        """
        Calculate Expected Information Gain (EIG).
        Proportional to novelty (distance) against the context.

        Returns value in [0.0, 1.0] where:
        - 0.0 = candidate is identical to context (no new information)
        - 1.0 = candidate is maximally different from context (max information)
        """
        return candidate.normalized_distance(context)

    async def _current_context_vector(self, sample_n: int = 50) -> BinaryHDV:
        """Superpose a slice of working memory (HOT tier) into a single context vector."""
        recent_nodes = await self.tier_manager.get_hot_recent(sample_n)

        if not recent_nodes:
            return BinaryHDV.zeros(self.dimension)

        vectors = [n.hdv for n in recent_nodes]
        if not vectors:
            return BinaryHDV.zeros(self.dimension)

        return majority_bundle(vectors)

    # ==========================================================================
    # Private Helper Methods for store() - Extracted for maintainability
    # ==========================================================================

    async def _encode_input(
        self,
        content: str,
        metadata: Optional[Dict[str, Any]] = None,
        goal_id: Optional[str] = None,
    ) -> Tuple[BinaryHDV, Dict[str, Any]]:
        """
        Encode input content to BinaryHDV and bind goal context if present.

        Args:
            content: The text content to encode.
            metadata: Optional metadata dictionary (will be mutated if goal_id present).
            goal_id: Optional goal identifier to bind as context.

        Returns:
            Tuple of (encoded BinaryHDV, updated metadata dict).
        """
        # Encode content (CPU bound operation)
        content_vec = await self._run_in_thread(self.binary_encoder.encode, content)

        # Initialize metadata if needed
        if metadata is None:
            metadata = {}

        final_vec = content_vec

        # Bind goal context if provided
        if goal_id:
            goal_vec = await self._run_in_thread(
                self.binary_encoder.encode, f"GOAL_CONTEXT_{goal_id}"
            )
            final_vec = content_vec.xor_bind(goal_vec)
            metadata['goal_context'] = goal_id

        return final_vec, metadata

    async def _evaluate_tier(
        self,
        encoded_vec: BinaryHDV,
        metadata: Dict[str, Any],
    ) -> Dict[str, Any]:
        """
        Calculate epistemic valuation (EIG) and update metadata accordingly.

        Args:
            encoded_vec: The encoded BinaryHDV to evaluate.
            metadata: Metadata dictionary to update with EIG values.

        Returns:
            Updated metadata dictionary with EIG information.
        """
        if self.epistemic_drive_active:
            ctx_vec = await self._current_context_vector(sample_n=50)
            eig = self.calculate_eig(encoded_vec, ctx_vec)
            metadata["eig"] = float(eig)

            if eig >= self.surprise_threshold:
                metadata.setdefault("tags", [])
                if isinstance(metadata["tags"], list):
                    metadata["tags"].append("epistemic_high")
        else:
            metadata.setdefault("eig", 0.0)

        return metadata

    async def _persist_memory(
        self,
        content: str,
        encoded_vec: BinaryHDV,
        metadata: Dict[str, Any],
    ) -> MemoryNode:
        """
        Create MemoryNode and persist to tier manager and disk.

        Phase 4.3: Automatically sets previous_id for episodic chaining.

        Args:
            content: Original text content.
            encoded_vec: Encoded BinaryHDV for the content.
            metadata: Metadata dictionary for the node.

        Returns:
            The created and persisted MemoryNode.
        """
        async with self._store_lock:
            previous_id = self._last_stored_id

            # Create node with unique ID
            node_id = str(uuid.uuid4())
            node = MemoryNode(
                id=node_id,
                hdv=encoded_vec,
                content=content,
                metadata=metadata,
                previous_id=previous_id,  # Phase 4.3: Episodic chaining
            )

            # Map EIG/Importance
            node.epistemic_value = float(metadata.get("eig", 0.0))
            node.calculate_ltp()

            # Store in Tier Manager (starts in HOT)
            await self.tier_manager.add_memory(node)

            # Append to persistence log (Legacy/Backup)
            await self._append_persisted(node)

            # Update linear episodic chain head only after successful persistence.
            self._last_stored_id = node.id

            return node

    async def _trigger_post_store(
        self,
        node: MemoryNode,
        metadata: Dict[str, Any],
    ) -> None:
        """
        Execute post-store triggers: subconscious queue and background dream.

        Gap-filled memories must NOT re-enter the dream/gap loop to prevent
        an indefinite store -> dream -> detect -> fill -> store cycle.

        Args:
            node: The MemoryNode that was stored.
            metadata: Metadata dictionary (checked for gap fill source).
        """
        _is_gap_fill = metadata.get("source") == "llm_gap_fill"

        # Phase 12.1: Aggressive Synapse Formation (Auto-bind).
        # Fix 4: collect all bindings first, persist synapses only once at the end.
        if hasattr(self.config, 'synapse') and self.config.synapse.auto_bind_on_store:
            similar_nodes = await self.query(
                node.content,
                top_k=3,
                associative_jump=False,
                track_gaps=False,
            )
            bind_pairs = [
                (node.id, neighbor_id)
                for neighbor_id, similarity in similar_nodes
                if neighbor_id != node.id
                and similarity >= self.config.synapse.similarity_threshold
            ]
            if bind_pairs:
                await self._auto_bind_batch(bind_pairs)

        self.subconscious_queue.append(node.id)

        if not _is_gap_fill:
            await self._background_dream(depth=1)

    # ==========================================================================
    # Main store() method - Orchestration only
    # ==========================================================================

    # Maximum allowed content length (Fix 5: input validation)
    _MAX_CONTENT_LENGTH: int = 100_000

    @timer(STORE_DURATION_SECONDS, labels={"tier": "hot"})
    @traced("store_memory")
    async def store(
        self,
        content: str,
        metadata: Optional[Dict[str, Any]] = None,
        goal_id: Optional[str] = None,
        project_id: Optional[str] = None,
    ) -> str:
        """
        Store new memory with holographic encoding.

        This method orchestrates the memory storage pipeline:
        1. Validate input
        2. Encode input content
        3. Evaluate tier placement via EIG
        4. Persist to storage
        5. Trigger post-store processing

        Args:
            content: The text content to store. Must be non-empty and ≤100 000 chars.
            metadata: Optional metadata dictionary.
            goal_id: Optional goal identifier for context binding.
            project_id: Optional project identifier for isolation masking (Phase 4.1).

        Returns:
            The unique identifier of the stored memory node.

        Raises:
            ValueError: If content is empty or exceeds the maximum allowed length.
            RuntimeError: If the engine has not been initialized via initialize().
        """
        # Fix 5: Input validation
        if not content or not content.strip():
            raise ValueError("Memory content cannot be empty or whitespace-only.")
        if len(content) > self._MAX_CONTENT_LENGTH:
            raise ValueError(
                f"Memory content is too long ({len(content):,} chars). "
                f"Maximum: {self._MAX_CONTENT_LENGTH:,}."
            )
        if not self._initialized:
            raise RuntimeError(
                "HAIMEngine.initialize() must be awaited before calling store()."
            )

        # 1. Encode input and bind goal context
        encoded_vec, updated_metadata = await self._encode_input(content, metadata, goal_id)

        # 1b. Apply project isolation mask (Phase 4.1)
        if project_id:
            encoded_vec = self.isolation_masker.apply_mask(encoded_vec, project_id)
            updated_metadata['project_id'] = project_id

        # 2. Calculate EIG and evaluate tier placement
        updated_metadata = await self._evaluate_tier(encoded_vec, updated_metadata)

        # 3. Create and persist memory node
        node = await self._persist_memory(content, encoded_vec, updated_metadata)

        # Phase 5.1: If agent_id in metadata, push to Working Memory and log Episode event
        agent_id = updated_metadata.get("agent_id")
        if agent_id:
            if self.working_memory:
                from .memory_model import WorkingMemoryItem
                self.working_memory.push_item(
                    agent_id, 
                    WorkingMemoryItem(
                        id=f"wm_{node.id[:8]}",
                        agent_id=agent_id,
                        created_at=datetime.utcnow(),
                        ttl_seconds=3600,
                        content=content,
                        kind="observation",
                        importance=node.epistemic_value or 0.5,
                        tags=updated_metadata.get("tags", []),
                        hdv=encoded_vec
                    )
                )
            
            episode_id = updated_metadata.get("episode_id")
            if episode_id and self.episodic_store:
                self.episodic_store.append_event(
                    episode_id=episode_id,
                    kind="observation",
                    content=content,
                    metadata=updated_metadata
                )

        # 4. Trigger post-store processing
        await self._trigger_post_store(node, updated_metadata)

        # 5. Update queue length metric
        update_queue_length(len(self.subconscious_queue))

        logger.info(f"Stored memory {node.id} (EIG: {updated_metadata.get('eig', 0.0):.4f})")
        return node.id

    async def delete_memory(self, node_id: str) -> bool:
        """
        Delete a memory from all internal states and storage tiers.
        Returns True if something was deleted.

        Phase 4.0: uses SynapseIndex.remove_node() for O(k) removal.
        """
        logger.info(f"Deleting memory {node_id}")

        # 1. Remove from TierManager (HOT/WARM/COLD-pending)
        deleted = await self.tier_manager.delete_memory(node_id)

        # 2. Remove from subconscious queue if present
        if node_id in self.subconscious_queue:
            self.subconscious_queue.remove(node_id)

        # 3. Phase 4.0: clean up via SynapseIndex (O(k)).
        # Fix 2: legacy dict rebuild removed — _synapse_index is authoritative.
        async with self.synapse_lock:
            removed_count = self._synapse_index.remove_node(node_id)

        if removed_count:
            await self._save_synapses()

        return deleted

    async def log_decision(self, context_text: str, outcome: float) -> None:
        """
        Phase 12.3: Logs a user decision or feedback context to update preference vector.
        Outcome should be positive (e.g. 1.0) or negative (e.g. -1.0).
        """
        vec = await self._run_in_thread(self.binary_encoder.encode, context_text)
        self.preference_store.log_decision(vec, outcome)

    async def close(self):
        """Perform graceful shutdown of engine components."""
        logger.info("Shutting down HAIMEngine...")

        # Phase 4.0: stop background workers
        if self._semantic_worker:
            await self._semantic_worker.stop()
        if self._immunology:
            await self._immunology.stop()
        if self._gap_filler:
            await self._gap_filler.stop()
        if self._subconscious_ai:
            await self._subconscious_ai.stop()

        await self._save_synapses()
        if self.tier_manager.use_qdrant and self.tier_manager.qdrant:
            await self.tier_manager.qdrant.close()

    @timer(QUERY_DURATION_SECONDS)
    @traced("query_memory")
    async def query(
        self,
        query_text: str,
        top_k: int = 5,
        associative_jump: bool = True,
        track_gaps: bool = True,
        project_id: Optional[str] = None,
        time_range: Optional[Tuple[datetime, datetime]] = None,
        chrono_weight: bool = True,
        chrono_lambda: float = 0.0001,
        include_neighbors: bool = False,
        metadata_filter: Optional[Dict[str, Any]] = None,
        include_cold: bool = False,
    ) -> List[Tuple[str, float]]:
        """
        Query memories using Hamming distance.
        Searches HOT tier and limited WARM tier.

        Phase 4.0 additions:
          - XOR attention masking re-ranks results for novelty.
          - Gap detection runs on low-confidence results (disabled when
            track_gaps=False to prevent dream-loop feedback).

        Phase 4.1 additions:
          - project_id applies isolation mask to query for project-scoped search.

        Phase 4.3 additions (Temporal Recall):
          - time_range: Filter to memories within (start, end) datetime range.
          - chrono_weight: Apply temporal decay to boost newer memories.
            Formula: Final_Score = Semantic_Similarity * (1 / (1 + lambda * Time_Delta))
          - chrono_lambda: Decay rate in seconds^-1 (default: 0.0001 ~ 2.7h half-life).
          - include_neighbors: Also fetch temporal neighbors (previous/next) for top results.
          - include_cold: Include COLD tier in the search (bounded linear scan, default False).

        Fix 3: Triggers anticipatory preloading (Phase 13.2) as fire-and-forget after returning.
        """
        # Encode Query
        query_vec = await self._run_in_thread(self.binary_encoder.encode, query_text)

        # Phase 12.2: Context Tracking
        is_shift, sim = self.topic_tracker.add_query(query_vec)
        if is_shift:
            logger.info(f"Context shifted during query. (sim {sim:.3f})")

        # Phase 4.1: Apply project isolation mask to query
        if project_id:
            query_vec = self.isolation_masker.apply_mask(query_vec, project_id)

        # 1. Primary Search (Accelerated FAISS/HNSW + Qdrant)
        # Phase 4.3: Pass time_range to tier_manager if filtering needed
        search_results = await self.tier_manager.search(
            query_vec,
            top_k=top_k * 2,
            time_range=time_range,
            metadata_filter=metadata_filter,
            include_cold=include_cold,
        )

        scores: Dict[str, float] = {}
        now_ts = datetime.now(timezone.utc).timestamp()
        mem_map: Dict[str, MemoryNode] = {}

        if chrono_weight and search_results:
            mems = await self.tier_manager.get_memories_batch(
                [nid for nid, _ in search_results]
            )
            mem_map = {m.id: m for m in mems if m}

        for nid, base_sim in search_results:
            # Boost by synaptic health (Phase 4.0: use SynapseIndex.boost for O(k))
            boost = self._synapse_index.boost(nid)
            score = base_sim * boost

            # Phase 4.3: Chrono-weighting (temporal decay)
            if chrono_weight and score > 0:
                mem = mem_map.get(nid)
                if mem:
                    time_delta = max(0.0, now_ts - mem.created_at.timestamp())  # seconds since creation
                    # Formula: Final = Semantic * (1 / (1 + lambda * time_delta))
                    decay_factor = 1.0 / (1.0 + chrono_lambda * time_delta)
                    score = score * decay_factor

            # Phase 12.3: Preference Learning Bias
            if self.preference_store.config.enabled and self.preference_store.preference_vector is not None:
                mem = mem_map.get(nid)
                if not mem:
                    mem = await self.tier_manager.get_memory(nid)
                    if mem and mem.id not in mem_map:
                        mem_map[mem.id] = mem
                if mem:
                    score = self.preference_store.bias_score(mem.hdv, score)

            scores[nid] = score

        # Phase 5.1: Boost context matching Working Memory 
        agent_id = metadata_filter.get("agent_id") if metadata_filter else None
        if agent_id and self.working_memory:
            wm_state = self.working_memory.get_state(agent_id)
            if wm_state:
                wm_texts = [item.content for item in wm_state.items]
                if wm_texts:
                    # Very lightweight lexical boost for items currently in working memory
                    q_lower = query_text.lower()
                    for nid in scores:
                        mem = mem_map.get(nid) # Assuming already cached from chrono weighting
                        if mem and mem.content:
                            if any(w_text.lower() in mem.content.lower() for w_text in wm_texts):
                                scores[nid] *= 1.15 # 15% boost for WM overlap

        # 2. Associative Spreading (via SynapseIndex for O(1) adjacency lookup)
        if associative_jump and self._synapse_index:
            top_seeds = sorted(scores.items(), key=lambda x: x[1], reverse=True)[:3]
            augmented_scores = scores.copy()

            for seed_id, seed_score in top_seeds:
                if seed_score <= 0:
                    continue

                neighbour_synapses = self._synapse_index.neighbours(seed_id)

                for syn in neighbour_synapses:
                    neighbor = (
                        syn.neuron_b_id if syn.neuron_a_id == seed_id else syn.neuron_a_id
                    )
                    if neighbor not in augmented_scores:
                        mem = await self.tier_manager.get_memory(neighbor)
                        if mem:
                            if metadata_filter:
                                match = True
                                node_meta = mem.metadata or {}
                                for k, v in metadata_filter.items():
                                    if node_meta.get(k) != v:
                                        match = False
                                        break
                                if not match:
                                    continue
                            augmented_scores[neighbor] = query_vec.similarity(mem.hdv)

                    if neighbor in augmented_scores:
                        spread = seed_score * syn.get_current_strength() * 0.3
                        augmented_scores[neighbor] += spread

            scores = augmented_scores

        # Phase 4.0: XOR attention re-ranking
        attention_mask = None
        top_results: List[Tuple[str, float]] = sorted(
            scores.items(), key=lambda x: x[1], reverse=True
        )[:top_k]

        if scores:
            # Build context key from recent HOT nodes
            recent_nodes = await self.tier_manager.get_hot_recent(
                self.attention_masker.config.context_sample_n
            )
            if recent_nodes:
                ctx_vecs = [n.hdv for n in recent_nodes]
                ctx_key = self.attention_masker.build_context_key(ctx_vecs)
                attention_mask = self.attention_masker.build_attention_mask(query_vec, ctx_key)

                # Collect HDVs for re-ranking (only HOT nodes available synchronously)
                mem_vecs: Dict[str, BinaryHDV] = {}
                async with self.tier_manager.lock:
                    for nid in list(scores.keys()):
                        node = self.tier_manager.hot.get(nid)
                        if node:
                            mem_vecs[nid] = node.hdv

                ranked = self.attention_masker.rerank(scores, mem_vecs, attention_mask)
                top_results = self.attention_masker.extract_scores(ranked)[:top_k]

        # Phase 4.0: Knowledge gap detection
        # Disabled during dream cycles to break the store->dream->gap->fill->store loop.
        if track_gaps:
            asyncio.ensure_future(
                self.gap_detector.assess_query(query_text, top_results, attention_mask)
            )

        # Phase 4.3: Sequential Context Window
        # Fetch temporal neighbors (previous_id chain and next in chain)
        if include_neighbors and top_results:
            neighbor_ids: set = set()
            for result_id, _ in top_results[:3]:  # Only for top 3 results
                mem = await self.tier_manager.get_memory(result_id)
                if not mem:
                    continue

                # Get the memory that came before this one (if episodic chain exists)
                if mem.previous_id:
                    prev_mem = await self.tier_manager.get_memory(mem.previous_id)
                    if prev_mem and prev_mem.id not in scores:
                        if metadata_filter:
                            match = True
                            p_meta = prev_mem.metadata or {}
                            for k, v in metadata_filter.items():
                                if p_meta.get(k) != v:
                                    match = False
                                    break
                            if not match:
                                continue
                        neighbor_ids.add(prev_mem.id)

                # Try to find the memory that follows this one (has this as previous_id).
                # Use the typed TierManager wrapper so we always work with MemoryNode,
                # not raw models.Record from Qdrant.
                next_mem = await self.tier_manager.get_next_in_chain(result_id)
                if next_mem and next_mem.id not in scores:
                    neighbor_ids.add(next_mem.id)

            # Add neighbors with their semantic scores (no chrono boost for context)
            for neighbor_id in neighbor_ids:
                mem = await self.tier_manager.get_memory(neighbor_id)
                if mem:
                    neighbor_score = query_vec.similarity(mem.hdv)
                    top_results.append((neighbor_id, neighbor_score * 0.8))  # Slightly discounted

            # Re-sort after adding neighbors, but preserve query() top_k contract.
            top_results = sorted(top_results, key=lambda x: x[1], reverse=True)[:top_k]

        # Phase 13.2 (Fix 3): Anticipatory preloading — fire-and-forget so it
        # never blocks the caller. Only activated when the engine is fully warm.
        if top_results and self._initialized and self.config.anticipatory.enabled:
            asyncio.ensure_future(
                self.anticipatory_engine.predict_and_preload(top_results[0][0])
            )

        return top_results

    async def get_context_nodes(self, top_k: int = 3) -> List[Tuple[str, float]]:
        """
        Phase 12.2: Contextual Awareness
        Retrieves the top_k most relevant nodes relating to the current topic context vector.
        Should be explicitly used by prompt builders before LLM logic injection.
        """
        if not self.topic_tracker.config.enabled:
            return []
            
        ctx = self.topic_tracker.get_context()
        if ctx is None:
            return []
            
        results = await self.tier_manager.search(
            ctx,
            top_k=top_k,
            time_range=None,
            metadata_filter=None,
        )
        return results
    async def _background_dream(self, depth: int = 2):
        """
        Passive Subconscious – strengthen synapses in idle cycles.

        Uses a semaphore so at most one dream task runs concurrently,
        and passes track_gaps=False so dream queries cannot feed the
        gap detector (breaking the store→dream→gap→fill→store loop).
        """
        if not self.subconscious_queue:
            return

        # Non-blocking: if a dream is already in progress, skip this cycle.
        if self._dream_sem.locked():
            return

        async with self._dream_sem:
            stim_id = self.subconscious_queue.popleft()
            stim_node = await self.tier_manager.get_memory(stim_id)
            if not stim_node:
                return

            potential_connections = await self.query(
                stim_node.content,
                top_k=depth + 1,
                associative_jump=False,
                track_gaps=False,   # ← no gap detection inside dream
            )

            for neighbor_id, similarity in potential_connections:
                if neighbor_id != stim_id and similarity > 0.15:
                    await self.bind_memories(stim_id, neighbor_id, success=True)

    async def orchestrate_orch_or(self, max_collapse: int = 3) -> List[MemoryNode]:
        """
        Collapse active HOT-tier superposition by a simple free-energy proxy.

        The score combines LTP (long-term stability), epistemic value (novelty),
        and access_count (usage evidence).
        """
        async with self.tier_manager.lock:
            active_nodes = list(self.tier_manager.hot.values())
        if not active_nodes or max_collapse <= 0:
            return []

        def score(node: MemoryNode) -> float:
            ltp = float(getattr(node, "ltp_strength", 0.0))
            epistemic = float(getattr(node, "epistemic_value", 0.0))
            access = float(getattr(node, "access_count", 0))
            return (0.6 * ltp) + (0.3 * epistemic) + (0.1 * np.log1p(access))

        return sorted(active_nodes, key=score, reverse=True)[:max_collapse]

    async def _auto_bind_batch(
        self,
        pairs: List[Tuple[str, str]],
        success: bool = True,
        weight: float = 1.0,
    ) -> None:
        """
        Fix 4: Bind multiple (id_a, id_b) pairs in one pass, saving synapses once.

        Used by auto-bind in _trigger_post_store() to avoid N disk writes per store.
        """
        async with self.synapse_lock:
            for id_a, id_b in pairs:
                mem_a = await self.tier_manager.get_memory(id_a)
                mem_b = await self.tier_manager.get_memory(id_b)
                if mem_a and mem_b:
                    self._synapse_index.add_or_fire(id_a, id_b, success=success, weight=weight)
        await self._save_synapses()

    async def bind_memories(self, id_a: str, id_b: str, success: bool = True, weight: float = 1.0):
        """
        Bind two memories by ID.

        Fix 2: delegates exclusively to SynapseIndex — legacy dict sync removed.
        The legacy self.synapses / self.synapse_adjacency attributes remain for
        backward compatibility but are only populated at startup from disk.
        """
        mem_a = await self.tier_manager.get_memory(id_a)
        mem_b = await self.tier_manager.get_memory(id_b)

        if not mem_a or not mem_b:
            return

        async with self.synapse_lock:
            self._synapse_index.add_or_fire(id_a, id_b, success=success, weight=weight)

        await self._save_synapses()

    async def get_node_boost(self, node_id: str) -> float:
        """
        Compute synaptic boost for scoring.

        Phase 4.0: O(k) via SynapseIndex (was O(k) before but with lock overhead).
        """
        return self._synapse_index.boost(node_id)

    async def cleanup_decay(self, threshold: float = 0.1):
        """
        Remove synapses that have decayed below the threshold.

        Phase 4.0: O(E) via SynapseIndex.compact(), no lock required for the index itself.
        Also syncs any legacy dict entries into the index before compacting.
        """
        async with self.synapse_lock:
            # Retain legacy→index sync so tests that write to self.synapses directly
            # still get their entries registered (Fix 2: sync only in this direction).
            for syn in list(self.synapses.values()):
                if self._synapse_index.get(syn.neuron_a_id, syn.neuron_b_id) is None:
                    self._synapse_index.register(syn)

            removed = self._synapse_index.compact(threshold)

        if removed:
            logger.info(f"cleanup_decay: pruned {removed} synapses below {threshold}")
            await self._save_synapses()

    async def get_stats(self) -> Dict[str, Any]:
        """Aggregate statistics from engine components."""
        tier_stats = await self.tier_manager.get_stats()

        async with self.synapse_lock:
            syn_count = len(self._synapse_index)

        stats = {
            "engine_version": "4.5.0",
            "dimension": self.dimension,
            "encoding": "binary_hdv",
            "tiers": tier_stats,
            "concepts_count": len(self.soul.concepts),
            "symbols_count": len(self.soul.symbols),
            "synapses_count": syn_count,
            "synapse_index": self._synapse_index.stats,
            "subconscious_backlog": len(self.subconscious_queue),
            # Phase 4.0
            "gap_detector": self.gap_detector.stats,
            "immunology": self._immunology.stats if self._immunology else {},
            "semantic_consolidation": (
                self._semantic_worker.stats if self._semantic_worker else {}
            ),
            # Phase 4.4: Subconscious AI worker stats (BETA)
            "subconscious_ai": (
                self._subconscious_ai.stats if self._subconscious_ai else {}
            ),
            # Phase 4.5: RecursiveSynthesizer stats
            "recursive_synthesizer": (
                self._recursive_synthesizer.stats if self._recursive_synthesizer else {}
            ),
            "timestamp": datetime.now(timezone.utc).isoformat(),
        }
        return stats

    def encode_content(self, content: str) -> BinaryHDV:
        """Encode text to Binary HDV."""
        return self.binary_encoder.encode(content)

    # ── Phase 4.0: Gap filling ─────────────────────────────────────

    async def enable_gap_filling(
        self,
        llm_integrator,
        config: Optional["GapFillerConfig"] = None,
    ) -> None:
        """
        Attach an LLM integrator to autonomously fill knowledge gaps.

        Args:
            llm_integrator: HAIMLLMIntegrator instance.
            config: Optional GapFillerConfig overrides.
        """
        if self._gap_filler:
            await self._gap_filler.stop()

        self._gap_filler = GapFiller(
            engine=self,
            llm_integrator=llm_integrator,
            gap_detector=self.gap_detector,
            config=config or GapFillerConfig(),
        )
        await self._gap_filler.start()
        logger.info("Phase 4.0 GapFiller started.")

    async def enable_recursive_synthesis(
        self,
        llm_call: Optional[Any] = None,
        config: Optional["SynthesizerConfig"] = None,
    ) -> None:
        """
        Enable Phase 4.5 Recursive Synthesis Engine.

        Args:
            llm_call: Optional callable for LLM-powered decomposition and synthesis.
                     Signature: (prompt: str) -> str.
            config: Optional SynthesizerConfig overrides.
        """
        self._recursive_synthesizer = RecursiveSynthesizer(
            engine=self,
            config=config or SynthesizerConfig(),
            llm_call=llm_call,
        )
        logger.info("Phase 4.5 RecursiveSynthesizer enabled.")

    async def record_retrieval_feedback(
        self,
        node_id: str,
        helpful: bool,
        eig_signal: float = 1.0,
    ) -> None:
        """
        Record whether a retrieved memory was useful.

        Phase 4.0: feeds the Bayesian LTP updater for the node.

        Args:
            node_id: The memory node that was retrieved.
            helpful: Was the retrieval actually useful?
            eig_signal: Strength of evidence (0–1).
        """
        node = await self.tier_manager.get_memory(node_id)
        if node:
            updater = get_bayesian_updater()
            updater.observe_node_retrieval(node, helpful=helpful, eig_signal=eig_signal)

    async def register_negative_feedback(self, query_text: str) -> None:
        """
        Signal that a recent query was not adequately answered.
        Creates a high-priority gap record for LLM gap-filling.
        """
        await self.gap_detector.register_negative_feedback(query_text)

    async def get_memory(self, node_id: str) -> Optional[MemoryNode]:
        """Retrieve memory via TierManager."""
        return await self.tier_manager.get_memory(node_id)

    # --- Legacy Helpers (for migration compatibility) ---

    def _legacy_encode_content_numpy(self, content: str) -> np.ndarray:
        """
        Original localized encoding logic for backward compatibility.
        Used only for migrating legacy data.
        """
        tokens = re.findall(r'\w+', content.lower())
        if not tokens:
            seed_bytes = hashlib.shake_256(content.encode()).digest(4)
            seed = int.from_bytes(seed_bytes, 'little')
            return np.random.RandomState(seed).choice([-1, 1], size=self.dimension)

        combined = np.zeros(self.dimension)
        for t in tokens:
            t_vec = self._get_token_vector(t, self.dimension)
            combined += t_vec

        v = np.sign(combined)
        v[v == 0] = np.random.RandomState(42).choice([-1, 1], size=np.sum(v == 0))
        return v.astype(int)

    async def _load_legacy_if_needed(self):
        """Load from memory.jsonl into TierManager, converting to BinaryHDV."""
        if not os.path.exists(self.persist_path):
            return

        logger.info(f"Loading legacy memory from {self.persist_path}")

        def _load():
            try:
                with open(self.persist_path, 'r', encoding='utf-8') as f:
                    return f.readlines()
            except Exception:
                return []

        lines = await self._run_in_thread(_load)

        for line in lines:
            line = line.strip()
            if not line:
                continue
            try:
                rec = json.loads(line)
                content = rec.get('content', '')
                if not content:
                    continue

                node_id = rec.get('id')

                # Always convert to BinaryHDV
                hdv = self.binary_encoder.encode(content)

                node = MemoryNode(
                    id=node_id,
                    hdv=hdv,
                    content=content,
                    metadata=rec.get('metadata') or {}
                )

                # Restore timestamps if available
                if 'created_at' in rec:
                    node.created_at = datetime.fromisoformat(rec['created_at'])

                # Phase 4.3: Restore episodic chain link
                if 'previous_id' in rec:
                    node.previous_id = rec['previous_id']

                # Add to TierManager
                await self.tier_manager.add_memory(node)

            except Exception as e:
                logger.warning(f"Failed to load record: {e}")

    async def _load_synapses(self):
        """
        Load synapses from disk.

        Phase 4.0: uses SynapseIndex.load_from_file() which restores Bayesian state.
        """
        if not os.path.exists(self.synapse_path):
            return

        def _load():
            self._synapse_index.load_from_file(self.synapse_path)

        # Fix 2: _synapse_index is authoritative — legacy dicts no longer rebuilt.
        await self._run_in_thread(_load)

    async def _save_synapses(self):
        """
        Save synapses to disk in JSONL format.

        Phase 4.0: uses SynapseIndex.save_to_file() which includes Bayesian state.
        A dedicated _write_lock serialises concurrent callers so the file is never
        written by two coroutines at the same time.  Does NOT acquire synapse_lock.
        """
        path_snapshot = self.synapse_path

        def _save():
            self._synapse_index.save_to_file(path_snapshot)

        async with self._write_lock:
            await self._run_in_thread(_save)

    async def _append_persisted(self, node: MemoryNode):
        """Append-only log with Phase 4.3 temporal metadata."""

        def _append():
            try:
                with open(self.persist_path, 'a', encoding='utf-8') as f:
                    rec = {
                        'id': node.id,
                        'content': node.content,
                        'metadata': node.metadata,
                        'created_at': node.created_at.isoformat(),
                        # Phase 4.3: Temporal metadata for indexing
                        'unix_timestamp': node.unix_timestamp,
                        'iso_date': node.iso_date,
                        'previous_id': node.previous_id,
                    }
                    f.write(json.dumps(rec) + "\n")
            except Exception as e:
                logger.error(f"Failed to persist memory: {e}")

        await self._run_in_thread(_append)

    async def persist_memory_snapshot(self, node: MemoryNode) -> None:
        """Persist a current snapshot of a memory node to the append-only log."""
        await self._append_persisted(node)

    # --- Conceptual Proxy ---

    async def define_concept(self, name: str, attributes: Dict[str, str]):
        await self._run_in_thread(self.soul.store_concept, name, attributes)

    async def reason_by_analogy(self, src: str, val: str, tgt: str):
        return await self._run_in_thread(self.soul.solve_analogy, src, val, tgt)

    async def cross_domain_inference(self, src: str, tgt: str, pat: str):
        return await self._run_in_thread(self.soul.solve_analogy, src, pat, tgt)

    async def inspect_concept(self, name: str, attr: str):
        return await self._run_in_thread(self.soul.extract_attribute, name, attr)