superlocalmemory 3.5.7 → 3.6.0

package/src/superlocalmemory/optimize/storage/db.py

@@ -0,0 +1,1016 @@
+"""CacheDB — wraps DatabaseManager for llmcache.db operations.
+
+INTERFACE-CONTRACT §1 CONFORMANCE: Every method below is the canonical API.
+No implementer may add aliases. Consumers (LLD-02/03/04/08) import this exact API.
+
+REUSE: DatabaseManager from superlocalmemory/src/superlocalmemory/storage/database.py
+  - WAL mode (database.py:66-74)
+  - busy_timeout=10000ms (database.py:41)
+  - _MAX_RETRIES=5 exponential backoff (database.py:127-148)
+  - _connect() with row_factory (database.py:95-100)
+  - transaction() context manager (database.py:102-116)
+  - execute() with retry (database.py:118-148)
+
+ENCRYPTION (resolves SEC-C-01 / CWE-312, NEW-M-01, NEW-M-02):
+  - All value BLOBs (llmcache_entries.value_blob) are AES-256-GCM encrypted.
+  - CCR original_blob is ALSO AES-256-GCM encrypted.
+  - Key derivation: PBKDF2-HMAC-SHA256(password=machine_id, salt=_per_db_salt, iter=100_000)
+  - Salt: os.urandom(32) generated ONCE at DB creation, stored in
+    llmcache_schema_version.description='salt:<hex>'. NO hardcoded salt.
+  - Nonce (12 bytes random) prepended to each ciphertext.
+  - llmcache.db file permissions set to 0o600 at creation.
+
+ISOLATION GUARANTEE: CacheDB MUST NOT import from superlocalmemory.storage.models.
+Enforced by test_no_memory_db_import() in test_db.py.
+
+FAIL-OPEN CONTRACT: Every public method catches sqlite3.Error, logs at WARNING,
+and returns a safe default (None, [], False, 0) rather than raising.
+"""
+
+from __future__ import annotations
+
+import dataclasses
+import json
+import logging
+import os
+import sqlite3
+import struct
+import time
+import uuid
+import zlib
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+from cryptography.hazmat.primitives.kdf.pbkdf2 import PBKDF2HMAC
+from cryptography.hazmat.primitives import hashes
+
+from superlocalmemory.storage.database import DatabaseManager
+from superlocalmemory.optimize.storage import schema as _schema
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_TENANT: str = "default"
+_ZLIB_LEVEL: int = 6
+_AES_NONCE_BYTES: int = 12
+_PBKDF2_ITERATIONS: int = 100_000
+
+LLMCACHE_DIRNAME: str = ".superlocalmemory"
+LLMCACHE_DBNAME: str = "llmcache.db"
+MID_FILENAME: str = ".llmcache_key"
+SALT_PREFIX: str = "salt:"
+
+_FORBIDDEN_MEMORY_TABLES: frozenset[str] = frozenset({
+    "memories", "atomic_facts", "profiles", "canonical_entities",
+    "entity_aliases", "consolidation_log", "trust_scores", "bm25_tokens",
+    "fact_retention", "core_memory_blocks", "ccq_consolidated_blocks",
+})
+
+
+# ---------------------------------------------------------------------------
+# MetricsSnapshot — single definition (INTERFACE-CONTRACT v2.2 §6)
+# ---------------------------------------------------------------------------
+
+@dataclass
+class MetricsSnapshot:
+    """Mirror of llmcache_metrics columns — names MUST match exactly."""
+    id: int = 1
+    hits: int = 0
+    misses: int = 0
+    calls_skipped: int = 0
+    tokens_saved_input: int = 0
+    tokens_saved_output: int = 0
+    tokens_saved_compress: int = 0
+    evictions: int = 0
+    latency_overhead_ms_sum: float = 0.0
+    latency_samples: int = 0
+    compress_runs: int = 0
+    compress_bytes_original: int = 0
+    compress_bytes_after: int = 0
+    cache_size_bytes: int = 0
+    cache_entry_count: int = 0
+    updated_at: float = 0.0
+
+    @property
+    def hit_rate(self) -> float:
+        total = self.hits + self.misses
+        return self.hits / total if total > 0 else 0.0
+
+    @property
+    def avg_latency_overhead_ms(self) -> float:
+        return (
+            self.latency_overhead_ms_sum / self.latency_samples
+            if self.latency_samples > 0 else 0.0
+        )
+
+    @property
+    def compression_ratio(self) -> float:
+        if self.compress_bytes_original == 0:
+            return 1.0
+        return self.compress_bytes_after / self.compress_bytes_original
+
+
+# ---------------------------------------------------------------------------
+# Supporting dataclasses
+# ---------------------------------------------------------------------------
+
+@dataclass
+class CacheRow:
+    """Return type of CacheDB.get() and get_by_id().
+
+    All fields are optional via default so that SELECT * on a table with
+    a different schema version does not blow up the constructor. The
+    caller is expected to read attributes they need; missing fields
+    return the default value.
+    """
+    entry_id: str = ""
+    cache_key: str = ""
+    tenant_id: str = "default"
+    model: str = ""
+    provider: str = ""
+    value: bytes = b""
+    created_at: str = ""
+    last_hit_at: str | None = None
+    ttl_expires: float | None = None
+    hit_count: int = 0
+    byte_size: int = 0
+    tag_json: str = "[]"
+    cache_tier: str = "exact"
+    compressed: int = 1
+    tags: list[str] = field(default_factory=list)
+
+
+@dataclass
+class BoundaryRow:
+    """Per-item vCache MLE boundary record."""
+    entry_id: str
+    logistic_t: float = 0.95
+    logistic_gamma: float = 10.0
+    sample_count: int = 0
+    updated_at: float = 0.0
+
+
+# ---------------------------------------------------------------------------
+# CacheDB
+# ---------------------------------------------------------------------------
+
+class CacheDB:
+    """SQLite cache store backed by DatabaseManager. Canonical API per INTERFACE-CONTRACT §1."""
+
+    _default_instance: "CacheDB | None" = None
+    _default_lock: Any = None  # lazy-import threading.Lock on first use
+
+    def __init__(self, db_path: Path | None = None) -> None:
+        if db_path is None:
+            db_path = Path.home() / LLMCACHE_DIRNAME / LLMCACHE_DBNAME
+        self._db_path = Path(db_path)
+        self._db_path.parent.mkdir(parents=True, exist_ok=True)
+        # If the file exists but is not a valid SQLite database (e.g. user
+        # accidentally pointed us at a binary blob), delete it and start fresh
+        # so that the proxy never gets wedged on init. The proxy is the
+        # critical path — we must be able to start even if the cache file
+        # is corrupted. The previous file is left in a `.corrupt` sidecar
+        # for forensic inspection.
+        if self._db_path.exists():
+            try:
+                import sqlite3 as _sq
+                test_conn = _sq.connect(str(self._db_path))
+                test_conn.execute("PRAGMA schema_version")
+                test_conn.close()
+            except Exception as exc:
+                corrupt_sidecar = self._db_path.with_suffix(
+                    self._db_path.suffix + ".corrupt"
+                )
+                try:
+                    if corrupt_sidecar.exists():
+                        corrupt_sidecar.unlink()
+                    self._db_path.replace(corrupt_sidecar)
+                    try:
+                        os.chmod(corrupt_sidecar, 0o600)
+                    except OSError:
+                        pass
+                    logger.error(
+                        "CacheDB: %s was not a SQLite file (%s) — moved to %s. "
+                        "Starting with a fresh cache.",
+                        self._db_path, exc, corrupt_sidecar,
+                    )
+                except OSError as move_exc:
+                    logger.error(
+                        "CacheDB: %s is not a SQLite file and could not be moved: %s",
+                        self._db_path, move_exc,
+                    )
+                    raise
+        self._db = DatabaseManager(self._db_path)
+        self._db.initialize(_schema)
+        # chmod 600 (SEC-C-01)
+        try:
+            os.chmod(self._db_path, 0o600)
+        except OSError as exc:
+            logger.warning("CacheDB: could not chmod 600 on %s: %s", self._db_path, exc)
+        self._salt = self._load_or_create_salt()
+        machine_id = self._get_machine_id()
+        self._aes_key = self._derive_aes_key(machine_id, self._salt)
+        self.assert_no_memory_db_tables()
+
+    # ---- context manager ----
+    def __enter__(self) -> "CacheDB":
+        return self
+
+    def __exit__(self, *args: Any) -> None:
+        self.close()
+
+    def close(self) -> None:
+        """No-op (per-call connection model)."""
+        self._db.close()
+
+    @property
+    def db_path(self) -> str:
+        return str(self._db_path)
+
+    # ---- encryption helpers ----
+
+    def _load_or_create_salt(self) -> bytes:
+        rows = self._db.execute(
+            "SELECT description FROM llmcache_schema_version "
+            "WHERE description LIKE 'salt:%' LIMIT 1"
+        )
+        if rows:
+            hex_salt = dict(rows[0])["description"][len(SALT_PREFIX):]
+            if len(hex_salt) != 64:
+                raise RuntimeError(
+                    f"llmcache.db: salt row is malformed (len={len(hex_salt)}) — DB may be corrupted."
+                )
+            return bytes.fromhex(hex_salt)
+        salt = os.urandom(32)
+        self._db.execute(
+            "INSERT INTO llmcache_schema_version (version, description) VALUES (?, ?)",
+            (1, f"{SALT_PREFIX}{salt.hex()}"),
+        )
+        return salt
+
+    def _get_machine_id(self) -> str:
+        """Return a stable machine identifier for AES key derivation."""
+        system = os.uname().sysname
+        mid: str | None = None
+        if system == "Darwin":
+            try:
+                import plistlib
+                import subprocess
+                out = subprocess.run(
+                    ["ioreg", "-rd1", "-c", "IOPlatformExpertDevice"],
+                    capture_output=True, timeout=2,
+                )
+                if out.returncode == 0:
+                    text = out.stdout.decode("utf-8", errors="ignore")
+                    for line in text.splitlines():
+                        if "IOPlatformUUID" in line and "=" in line:
+                            mid = line.split("=", 1)[1].strip().strip('"')
+                            if mid:
+                                break
+            except Exception:
+                mid = None
+        elif system == "Linux":
+            try:
+                mid = Path("/etc/machine-id").read_text(encoding="utf-8").strip()
+            except OSError:
+                mid = None
+        if not mid:
+            mid_file = Path.home() / LLMCACHE_DIRNAME / MID_FILENAME
+            if mid_file.exists():
+                try:
+                    mid = mid_file.read_text(encoding="utf-8").strip()
+                except OSError:
+                    mid = None
+            if not mid:
+                mid = uuid.uuid4().hex
+                try:
+                    mid_file.parent.mkdir(parents=True, exist_ok=True)
+                    mid_file.write_text(mid, encoding="utf-8")
+                    os.chmod(mid_file, 0o600)
+                except OSError as exc:
+                    logger.warning("CacheDB: could not persist machine id: %s", exc)
+        return mid
+
+    def _derive_aes_key(self, machine_id: str, salt: bytes) -> bytes:
+        kdf = PBKDF2HMAC(
+            algorithm=hashes.SHA256(),
+            length=32,
+            salt=salt,
+            iterations=_PBKDF2_ITERATIONS,
+        )
+        return kdf.derive(machine_id.encode("utf-8"))
+
+    def _encrypt(self, plaintext: bytes) -> bytes:
+        nonce = os.urandom(_AES_NONCE_BYTES)
+        aesgcm = AESGCM(self._aes_key)
+        ciphertext = aesgcm.encrypt(nonce, plaintext, associated_data=None)
+        return nonce + ciphertext
+
+    def _decrypt(self, blob: bytes) -> bytes:
+        if len(blob) <= _AES_NONCE_BYTES:
+            raise ValueError("Encrypted blob too short")
+        nonce = blob[:_AES_NONCE_BYTES]
+        ciphertext = blob[_AES_NONCE_BYTES:]
+        aesgcm = AESGCM(self._aes_key)
+        return aesgcm.decrypt(nonce, ciphertext, associated_data=None)
+
+    # ---- assertion ----
+
+    def assert_no_memory_db_tables(self) -> None:
+        """Raise RuntimeError if memory.db tables are present."""
+        rows = self._db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table'"
+        )
+        names = {dict(r)["name"] for r in rows}
+        found = _FORBIDDEN_MEMORY_TABLES & names
+        if found:
+            raise RuntimeError(
+                f"ISOLATION VIOLATION: llmcache.db contains memory.db tables: {sorted(found)}. "
+                "Wrong DB file opened. Aborting."
+            )
+
+    # ---- exact CRUD ----
+
+    def get(self, key: str, tenant_id: str) -> CacheRow | None:
+        try:
+            rows = self._db.execute(
+                "SELECT * FROM llmcache_entries "
+                "WHERE cache_key = ? AND tenant_id = ? "
+                "AND (ttl_expires IS NULL OR ttl_expires > ?) "
+                "LIMIT 1",
+                (key, tenant_id, time.time()),
+            )
+            if not rows:
+                return None
+            row = dict(rows[0])
+            # increment hit stats
+            self._db.execute(
+                "UPDATE llmcache_entries SET hit_count = hit_count + 1, "
+                "last_hit_at = strftime('%Y-%m-%dT%H:%M:%fZ', 'now') "
+                "WHERE cache_key = ? AND tenant_id = ?",
+                (key, tenant_id),
+            )
+            row["hit_count"] = row["hit_count"] + 1
+            # decrypt + decompress value
+            try:
+                plaintext = self._decrypt(row["value_blob"])
+                if row.get("compressed", 0):
+                    plaintext = zlib.decompress(plaintext)
+                row["value"] = plaintext
+            except (ValueError, zlib.error) as exc:
+                logger.warning("CacheDB.get decrypt failed (cache miss): %s", exc)
+                return None
+            return self._row_to_cacherow(row)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.get failed (cache miss): %s", exc)
+            return None
+
+    def set(
+        self,
+        key: str,
+        tenant_id: str,
+        value: bytes,
+        *,
+        model: str,
+        ttl_expires: float | None,
+        tags: list[str],
+    ) -> None:
+        try:
+            # 1) zlib-compress
+            compressed_blob = zlib.compress(value, level=_ZLIB_LEVEL)
+            # 2) AES-256-GCM encrypt
+            encrypted_blob = self._encrypt(compressed_blob)
+            entry_id = uuid.uuid4().hex
+            self._db.execute(
+                "INSERT INTO llmcache_entries "
+                "(entry_id, cache_key, tenant_id, model, provider, value_blob, compressed, "
+                " ttl_expires, tag_json, byte_size) "
+                "VALUES (?, ?, ?, ?, ?, ?, 1, ?, ?, ?) "
+                "ON CONFLICT(cache_key, tenant_id) DO UPDATE SET "
+                "  value_blob=excluded.value_blob, model=excluded.model, provider=excluded.provider, "
+                "  ttl_expires=excluded.ttl_expires, tag_json=excluded.tag_json, "
+                "  byte_size=excluded.byte_size, hit_count=0, last_hit_at=NULL, "
+                "  created_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')",
+                (
+                    entry_id, key, tenant_id, model, "",
+                    encrypted_blob, ttl_expires,
+                    json.dumps(tags or []), len(encrypted_blob),
+                ),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.set failed (fail-open): %s", exc)
+
+    def set_with_entry_id(
+        self,
+        key: str,
+        tenant_id: str,
+        value: bytes,
+        entry_id: str,
+        ttl_seconds: int | None = None,
+        tags: list[str] | None = None,
+    ) -> None:
+        """Like set(), but writes a caller-supplied entry_id instead of auto-generating one.
+
+        Used by semantic layer (index_entry) which pre-computes the entry_id from
+        the embedding vector to enable O(1) semantic-to-cache cross-reference.
+        """
+        try:
+            compressed_blob = zlib.compress(value, level=_ZLIB_LEVEL)
+            encrypted_blob = self._encrypt(compressed_blob)
+            ttl_expires = (time.time() + ttl_seconds) if ttl_seconds else None
+            self._db.execute(
+                "INSERT INTO llmcache_entries "
+                "(entry_id, cache_key, tenant_id, model, provider, value_blob, compressed, "
+                " ttl_expires, tag_json, byte_size) "
+                "VALUES (?, ?, ?, ?, ?, ?, 1, ?, ?, ?) "
+                "ON CONFLICT(cache_key, tenant_id) DO UPDATE SET "
+                "  entry_id=excluded.entry_id, value_blob=excluded.value_blob, "
+                "  model=excluded.model, provider=excluded.provider, "
+                "  ttl_expires=excluded.ttl_expires, tag_json=excluded.tag_json, "
+                "  byte_size=excluded.byte_size, hit_count=0, last_hit_at=NULL, "
+                "  created_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')",
+                (
+                    entry_id, key, tenant_id, "", "",
+                    encrypted_blob, ttl_expires,
+                    json.dumps(tags or []), len(encrypted_blob),
+                ),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.set_with_entry_id failed (fail-open): %s", exc)
+
+    def get_by_id(self, entry_id: str) -> CacheRow | None:
+        try:
+            rows = self._db.execute(
+                "SELECT * FROM llmcache_entries WHERE entry_id = ? LIMIT 1",
+                (entry_id,),
+            )
+            if not rows:
+                return None
+            row = dict(rows[0])
+            try:
+                plaintext = self._decrypt(row["value_blob"])
+                if row.get("compressed", 0):
+                    plaintext = zlib.decompress(plaintext)
+                row["value"] = plaintext
+            except (ValueError, zlib.error) as exc:
+                logger.warning("CacheDB.get_by_id decrypt failed: %s", exc)
+                return None
+            return self._row_to_cacherow(row)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.get_by_id failed: %s", exc)
+            return None
+
+    def delete(self, key: str, tenant_id: str) -> None:
+        try:
+            with self._db.transaction():
+                rows = self._db.execute(
+                    "SELECT entry_id FROM llmcache_entries WHERE cache_key = ? AND tenant_id = ?",
+                    (key, tenant_id),
+                )
+                entry_ids = [dict(r)["entry_id"] for r in rows]
+                if entry_ids:
+                    placeholders = ",".join("?" for _ in entry_ids)
+                    self._db.execute(
+                        f"DELETE FROM llmcache_semantic_vectors "
+                        f"WHERE entry_id IN ({placeholders})",
+                        tuple(entry_ids),
+                    )
+                self._db.execute(
+                    "DELETE FROM llmcache_tags WHERE cache_key = ? AND tenant_id = ?",
+                    (key, tenant_id),
+                )
+                self._db.execute(
+                    "DELETE FROM llmcache_entries WHERE cache_key = ? AND tenant_id = ?",
+                    (key, tenant_id),
+                )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.delete failed (fail-open): %s", exc)
+
+    def delete_batch(self, keys: list[tuple[str, str]]) -> int:
+        if not keys:
+            return 0
+        try:
+            deleted = 0
+            with self._db.transaction():
+                for key, tenant_id in keys:
+                    rows = self._db.execute(
+                        "SELECT entry_id FROM llmcache_entries WHERE cache_key = ? AND tenant_id = ?",
+                        (key, tenant_id),
+                    )
+                    entry_ids = [dict(r)["entry_id"] for r in rows]
+                    if entry_ids:
+                        placeholders = ",".join("?" for _ in entry_ids)
+                        self._db.execute(
+                            f"DELETE FROM llmcache_semantic_vectors "
+                            f"WHERE entry_id IN ({placeholders})",
+                            tuple(entry_ids),
+                        )
+                    self._db.execute(
+                        "DELETE FROM llmcache_tags WHERE cache_key = ? AND tenant_id = ?",
+                        (key, tenant_id),
+                    )
+                    self._db.execute(
+                        "DELETE FROM llmcache_entries WHERE cache_key = ? AND tenant_id = ?",
+                        (key, tenant_id),
+                    )
+                    deleted += 1
+            return deleted
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.delete_batch failed: %s", exc)
+            return 0
+
+    def sweep_expired(self, now: float) -> int:
+        try:
+            with self._db.transaction():
+                rows = self._db.execute(
+                    "SELECT entry_id FROM llmcache_entries "
+                    "WHERE ttl_expires IS NOT NULL AND ttl_expires < ?",
+                    (now,),
+                )
+                entry_ids = [dict(r)["entry_id"] for r in rows]
+                entries_deleted = len(entry_ids)
+                if entry_ids:
+                    placeholders = ",".join("?" for _ in entry_ids)
+                    self._db.execute(
+                        f"DELETE FROM llmcache_semantic_vectors "
+                        f"WHERE entry_id IN ({placeholders})",
+                        tuple(entry_ids),
+                    )
+                    self._db.execute(
+                        f"DELETE FROM llmcache_tags "
+                        f"WHERE cache_key IN ("
+                        f"  SELECT cache_key FROM llmcache_entries WHERE entry_id IN ({placeholders})"
+                        f")",
+                        tuple(entry_ids),
+                    )
+                self._db.execute(
+                    "DELETE FROM llmcache_entries "
+                    "WHERE ttl_expires IS NOT NULL AND ttl_expires < ?",
+                    (now,),
+                )
+                ccr_row = self._db.execute(
+                    "SELECT COUNT(*) AS n FROM llmcache_ccr_originals "
+                    "WHERE ttl_expires IS NOT NULL AND ttl_expires < ?",
+                    (now,),
+                )
+                ccr_count = int(dict(ccr_row[0])["n"]) if ccr_row else 0
+                self._db.execute(
+                    "DELETE FROM llmcache_ccr_originals "
+                    "WHERE ttl_expires IS NOT NULL AND ttl_expires < ?",
+                    (now,),
+                )
+            return entries_deleted + ccr_count
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.sweep_expired failed: %s", exc)
+            return 0
+
+    # ---- tags ----
+
+    def tag_register(self, key: str, tenant_id: str, tags: list[str]) -> None:
+        if not tags:
+            return
+        try:
+            with self._db.transaction():
+                for tag in tags:
+                    self._db.execute(
+                        "INSERT OR IGNORE INTO llmcache_tags (tag, cache_key, tenant_id) "
+                        "VALUES (?, ?, ?)",
+                        (tag, key, tenant_id),
+                    )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.tag_register failed: %s", exc)
+
+    def tag_keys(self, tag: str) -> list[tuple[str, str]]:
+        try:
+            rows = self._db.execute(
+                "SELECT cache_key, tenant_id FROM llmcache_tags WHERE tag = ?",
+                (tag,),
+            )
+            return [(dict(r)["cache_key"], dict(r)["tenant_id"]) for r in rows]
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.tag_keys failed: %s", exc)
+            return []
+
+    def invalidate_by_tag(self, tag: str) -> int:
+        try:
+            with self._db.transaction():
+                rows = self._db.execute(
+                    "SELECT DISTINCT e.entry_id FROM llmcache_entries e "
+                    "JOIN llmcache_tags t ON t.cache_key = e.cache_key AND t.tenant_id = e.tenant_id "
+                    "WHERE t.tag = ?",
+                    (tag,),
+                )
+                entry_ids = [dict(r)["entry_id"] for r in rows]
+                if not entry_ids:
+                    self._db.execute("DELETE FROM llmcache_tags WHERE tag = ?", (tag,))
+                    return 0
+                placeholders = ",".join("?" for _ in entry_ids)
+                self._db.execute(
+                    f"DELETE FROM llmcache_semantic_vectors WHERE entry_id IN ({placeholders})",
+                    tuple(entry_ids),
+                )
+                self._db.execute(
+                    f"DELETE FROM llmcache_entries WHERE entry_id IN ({placeholders})",
+                    tuple(entry_ids),
+                )
+                self._db.execute("DELETE FROM llmcache_tags WHERE tag = ?", (tag,))
+                return len(entry_ids)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.invalidate_by_tag failed: %s", exc)
+            return 0
+
+    # ---- semantic vectors ----
+
+    def vec_add(self, entry_id: str, tenant_id: str, vector: bytes, meta: dict) -> None:
+        try:
+            dim = int(meta.get("dim", len(vector) // 4))
+            model_name = str(meta.get("model", "nomic-ai/nomic-embed-text-v1.5"))
+            self._db.execute(
+                "INSERT OR REPLACE INTO llmcache_semantic_vectors "
+                "(entry_id, tenant_id, vector_blob, vector_dim, model_name) "
+                "VALUES (?, ?, ?, ?, ?)",
+                (entry_id, tenant_id, vector, dim, model_name),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.vec_add failed: %s", exc)
+
+    def vec_search(self, tenant_id: str, vector: bytes, top_k: int) -> list[tuple[str, float]]:
+        try:
+            rows = self._db.execute(
+                "SELECT entry_id, vector_blob FROM llmcache_semantic_vectors "
+                "WHERE tenant_id = ?",
+                (tenant_id,),
+            )
+            try:
+                import numpy as _np
+            except ImportError:
+                return []
+            q = _np.frombuffer(vector, dtype=_np.float32)
+            q_norm = float(_np.linalg.norm(q))
+            if q_norm == 0:
+                return []
+            scored: list[tuple[str, float]] = []
+            for r in rows:
+                rd = dict(r)
+                v = _np.frombuffer(rd["vector_blob"], dtype=_np.float32)
+                v_norm = float(_np.linalg.norm(v))
+                if v_norm == 0:
+                    continue
+                cos = float(_np.dot(q, v) / (q_norm * v_norm))
+                scored.append((rd["entry_id"], cos))
+            scored.sort(key=lambda t: t[1], reverse=True)
+            return scored[: max(0, top_k)]
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.vec_search failed: %s", exc)
+            return []
+
+    def vec_delete(self, entry_id: str) -> None:
+        try:
+            self._db.execute(
+                "DELETE FROM llmcache_semantic_vectors WHERE entry_id = ?",
+                (entry_id,),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.vec_delete failed: %s", exc)
+
+    # ---- vCache boundary + centroids (Phase 3 hooks, stub-implemented) ----
+
+    def boundary_get(self, entry_id: str) -> BoundaryRow | None:
+        try:
+            rows = self._db.execute(
+                "SELECT * FROM llmcache_boundaries WHERE entry_id = ? LIMIT 1",
+                (entry_id,),
+            )
+            if not rows:
+                return None
+            d = dict(rows[0])
+            return BoundaryRow(**d)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.boundary_get failed: %s", exc)
+            return None
+
+    def boundary_upsert(self, entry_id: str, row: BoundaryRow) -> None:
+        try:
+            self._db.execute(
+                "INSERT INTO llmcache_boundaries "
+                "(entry_id, logistic_t, logistic_gamma, sample_count, updated_at) "
+                "VALUES (?, ?, ?, ?, ?) "
+                "ON CONFLICT(entry_id) DO UPDATE SET "
+                "  logistic_t=excluded.logistic_t, logistic_gamma=excluded.logistic_gamma, "
+                "  sample_count=excluded.sample_count, updated_at=excluded.updated_at",
+                (
+                    entry_id, row.logistic_t, row.logistic_gamma,
+                    row.sample_count, row.updated_at,
+                ),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.boundary_upsert failed: %s", exc)
+
+    def centroid_get(self, tenant_id: str) -> bytes | None:
+        try:
+            rows = self._db.execute(
+                "SELECT centroid_blob FROM llmcache_centroids WHERE tenant_id = ? LIMIT 1",
+                (tenant_id,),
+            )
+            if not rows:
+                return None
+            return dict(rows[0])["centroid_blob"]
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.centroid_get failed: %s", exc)
+            return None
+
+    def centroid_update(self, tenant_id: str, centroid: bytes, n: int) -> None:
+        try:
+            self._db.execute(
+                "INSERT INTO llmcache_centroids (tenant_id, centroid_blob, n) "
+                "VALUES (?, ?, ?) "
+                "ON CONFLICT(tenant_id) DO UPDATE SET "
+                "  centroid_blob=excluded.centroid_blob, n=excluded.n, "
+                "  updated_at=strftime('%Y-%m-%dT%H:%M:%fZ','now')",
+                (tenant_id, centroid, n),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.centroid_update failed: %s", exc)
+
+    # ---- v2.2 supplementary boundary methods (LLD-03 Phase 3) ----
+
+    def get_all_boundaries(self) -> list[dict[str, Any]]:
+        """Return all vCache MLE boundary records (warm-start / rebuild).
+
+        LLD-03 §4.4: called at VCacheSemantic startup to populate the
+        in-memory boundary cache.
+
+        Returns:
+            List of dicts with keys: entry_id, t_hat, gamma_hat,
+            sample_count, updated_at. [] on error.
+        """
+        try:
+            rows = self._db.execute(
+                "SELECT entry_id, logistic_t, logistic_gamma, sample_count, updated_at "
+                "FROM llmcache_boundaries"
+            )
+            return [dict(r) for r in rows]
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.get_all_boundaries failed: %s", exc)
+            return []
+
+    def delete_boundary(self, entry_id: str) -> None:
+        """Delete a vCache boundary record. Fail-open."""
+        try:
+            self._db.execute(
+                "DELETE FROM llmcache_boundaries WHERE entry_id = ?",
+                (entry_id,),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.delete_boundary failed: %s", exc)
+
+    def get_entry_by_id(self, entry_id: str) -> dict[str, Any] | None:
+        """Fetch a cache entry by entry_id and return its decoded response dict.
+
+        LLD-03 §4.5 _fetch_response: needed to retrieve the response for a
+        semantically matched entry_id. Decrypts + decompresses the value_blob.
+
+        Returns:
+            Response dict (parsed JSON), or None on miss / error.
+        """
+        try:
+            rows = self._db.execute(
+                "SELECT value_blob, compressed FROM llmcache_entries "
+                "WHERE entry_id = ? LIMIT 1",
+                (entry_id,),
+            )
+            if not rows:
+                return None
+            row = dict(rows[0])
+            try:
+                plaintext = self._decrypt(row["value_blob"])
+                if row.get("compressed", 0):
+                    plaintext = zlib.decompress(plaintext)
+            except (ValueError, zlib.error) as exc:
+                logger.warning("CacheDB.get_entry_by_id decrypt failed: %s", exc)
+                return None
+            try:
+                import json as _json
+                return _json.loads(plaintext.decode("utf-8"))
+            except (UnicodeDecodeError, json.JSONDecodeError):
+                return {"raw_bytes": plaintext.hex()}
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.get_entry_by_id failed: %s", exc)
+            return None
+
+    # ---- CCR originals ----
+
+    def ccr_put(
+        self,
+        ccr_id: str,
+        original: bytes,
+        ttl_expires: float | None = None,
+    ) -> None:
+        import hashlib
+        try:
+            compressed = zlib.compress(original, level=_ZLIB_LEVEL)
+            encrypted = self._encrypt(compressed)
+            self._db.execute(
+                "INSERT OR REPLACE INTO llmcache_ccr_originals "
+                "(ccr_id, original_blob, compressed_hash, byte_size_orig, byte_size_comp, ttl_expires) "
+                "VALUES (?, ?, ?, ?, ?, ?)",
+                (
+                    ccr_id, encrypted,
+                    hashlib.sha256(original).hexdigest(),
+                    len(original), len(compressed),
+                    ttl_expires,
+                ),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.ccr_put failed: %s", exc)
+
+    def ccr_get(self, ccr_id: str) -> bytes | None:
+        try:
+            rows = self._db.execute(
+                "SELECT original_blob FROM llmcache_ccr_originals "
+                "WHERE ccr_id = ? "
+                "AND (ttl_expires IS NULL OR ttl_expires > ?)",
+                (ccr_id, time.time()),
+            )
+            if not rows:
+                return None
+            blob = dict(rows[0])["original_blob"]
+            plaintext = self._decrypt(blob)
+            return zlib.decompress(plaintext)
+        except (sqlite3.Error, ValueError, zlib.error) as exc:
+            logger.warning("CacheDB.ccr_get failed: %s", exc)
+            return None
+
+    def ccr_update_compressed(self, ccr_id: str, compressed: bytes) -> None:
+        import hashlib
+        try:
+            self._db.execute(
+                "UPDATE llmcache_ccr_originals "
+                "SET compressed_hash = ?, byte_size_comp = ? WHERE ccr_id = ?",
+                (hashlib.sha256(compressed).hexdigest(), len(compressed), ccr_id),
+            )
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.ccr_update_compressed failed: %s", exc)
+
+    # ---- v2 additions ----
+
+    def get_all_vectors(self, tenant_id: str) -> list[tuple[str, bytes]]:
+        try:
+            rows = self._db.execute(
+                "SELECT entry_id, vector_blob FROM llmcache_semantic_vectors "
+                "WHERE tenant_id = ?",
+                (tenant_id,),
+            )
+            return [(dict(r)["entry_id"], dict(r)["vector_blob"]) for r in rows]
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.get_all_vectors failed: %s", exc)
+            return []
+
+    # ---- metrics ----
+
+    def metrics_load(self) -> MetricsSnapshot:
+        try:
+            rows = self._db.execute("SELECT * FROM llmcache_metrics WHERE id = 1")
+            if not rows:
+                return MetricsSnapshot()
+            d = dict(rows[0])
+            return MetricsSnapshot(**d)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.metrics_load failed: %s", exc)
+            return MetricsSnapshot()
+
+    def metrics_flush(self, snap: MetricsSnapshot) -> None:
+        try:
+            params = dataclasses.asdict(snap)
+            sql = (
+                "INSERT INTO llmcache_metrics ("
+                "  id, hits, misses, calls_skipped, "
+                "  tokens_saved_input, tokens_saved_output, tokens_saved_compress, "
+                "  evictions, latency_overhead_ms_sum, latency_samples, "
+                "  compress_runs, compress_bytes_original, compress_bytes_after, "
+                "  cache_size_bytes, cache_entry_count, updated_at) "
+                "VALUES ("
+                "  :id, :hits, :misses, :calls_skipped, "
+                "  :tokens_saved_input, :tokens_saved_output, :tokens_saved_compress, "
+                "  :evictions, :latency_overhead_ms_sum, :latency_samples, "
+                "  :compress_runs, :compress_bytes_original, :compress_bytes_after, "
+                "  :cache_size_bytes, :cache_entry_count, :updated_at) "
+                "ON CONFLICT(id) DO UPDATE SET "
+                "  hits=excluded.hits, misses=excluded.misses, calls_skipped=excluded.calls_skipped, "
+                "  tokens_saved_input=excluded.tokens_saved_input, "
+                "  tokens_saved_output=excluded.tokens_saved_output, "
+                "  tokens_saved_compress=excluded.tokens_saved_compress, "
+                "  evictions=excluded.evictions, "
+                "  latency_overhead_ms_sum=excluded.latency_overhead_ms_sum, "
+                "  latency_samples=excluded.latency_samples, "
+                "  compress_runs=excluded.compress_runs, "
+                "  compress_bytes_original=excluded.compress_bytes_original, "
+                "  compress_bytes_after=excluded.compress_bytes_after, "
+                "  cache_size_bytes=excluded.cache_size_bytes, "
+                "  cache_entry_count=excluded.cache_entry_count, "
+                "  updated_at=excluded.updated_at"
+            )
+            with self._db.transaction():
+                self._db.execute(sql, params)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.metrics_flush failed: %s", exc)
+
+    # ---- convenience / non-contract helpers ----
+
+    def entry_exists(self, cache_key: str, tenant_id: str = _DEFAULT_TENANT) -> bool:
+        try:
+            rows = self._db.execute(
+                "SELECT 1 FROM llmcache_entries WHERE cache_key = ? AND tenant_id = ? LIMIT 1",
+                (cache_key, tenant_id),
+            )
+            return bool(rows)
+        except sqlite3.Error:
+            return False
+
+    def clear_tenant(self, tenant_id: str) -> int:
+        try:
+            with self._db.transaction():
+                rows = self._db.execute(
+                    "SELECT entry_id FROM llmcache_entries WHERE tenant_id = ?",
+                    (tenant_id,),
+                )
+                entry_ids = [dict(r)["entry_id"] for r in rows]
+                if entry_ids:
+                    placeholders = ",".join("?" for _ in entry_ids)
+                    self._db.execute(
+                        f"DELETE FROM llmcache_semantic_vectors "
+                        f"WHERE entry_id IN ({placeholders})",
+                        tuple(entry_ids),
+                    )
+                self._db.execute(
+                    "DELETE FROM llmcache_tags WHERE tenant_id = ?", (tenant_id,),
+                )
+                self._db.execute(
+                    "DELETE FROM llmcache_entries WHERE tenant_id = ?", (tenant_id,),
+                )
+                return len(entry_ids)
+        except sqlite3.Error as exc:
+            logger.warning("CacheDB.clear_tenant failed: %s", exc)
+            return 0
+
+    def entry_count(self, tenant_id: str = _DEFAULT_TENANT) -> int:
+        try:
+            rows = self._db.execute(
+                "SELECT COUNT(*) AS n FROM llmcache_entries "
+                "WHERE tenant_id = ? AND (ttl_expires IS NULL OR ttl_expires > ?)",
+                (tenant_id, time.time()),
+            )
+            return int(dict(rows[0])["n"]) if rows else 0
+        except sqlite3.Error:
+            return 0
+
+    def db_size_bytes(self) -> int:
+        try:
+            rows = self._db.execute("PRAGMA page_count")
+            page_count = int(dict(rows[0])["page_count"]) if rows else 0
+            rows = self._db.execute("PRAGMA page_size")
+            page_size = int(dict(rows[0])["page_size"]) if rows else 0
+            return page_count * page_size
+        except sqlite3.Error:
+            return 0
+
+    # ---- internal helpers ----
+
+    def _row_to_cacherow(self, row: dict) -> "CacheRow":
+        """Build a CacheRow from a sqlite3.Row dict, robust to extra/missing keys."""
+        kwargs: dict[str, Any] = {}
+        for fname in CacheRow.__dataclass_fields__:
+            if fname == "value":
+                continue
+            if fname in row:
+                kwargs[fname] = row[fname]
+        kwargs["value"] = row.get("value", b"")
+        try:
+            kwargs["tags"] = json.loads(row.get("tag_json") or "[]")
+        except (json.JSONDecodeError, TypeError):
+            kwargs["tags"] = []
+        return CacheRow(**kwargs)
+
+    # ---- singleton (INTERFACE-CONTRACT v2.2 §1) ----
+
+    @classmethod
+    def get_default(cls) -> "CacheDB":
+        if cls._default_instance is None:
+            import threading as _t
+            if cls._default_lock is None:
+                cls._default_lock = _t.Lock()
+            with cls._default_lock:
+                if cls._default_instance is None:
+                    cls._default_instance = cls()
+        return cls._default_instance
+
+    @classmethod
+    def reset_default(cls) -> None:
+        """Reset the singleton (testing only)."""
+        if cls._default_instance is not None:
+            try:
+                cls._default_instance.close()
+            except Exception:
+                pass
+        cls._default_instance = None