superlocalmemory 3.3.6 → 3.3.8

package/README.md

@@ -3,7 +3,8 @@
 </p>
 
 <h1 align="center">SuperLocalMemory V3.3</h1>
-<p align="center"><strong>The first local-only AI memory to break 74% retrieval on LoCoMo.<br/>No cloud. No APIs. No data leaves your machine.</strong></p>
+<p align="center"><strong>Every other AI forgets. Yours won't.</strong><br/><em>Infinite memory for Claude Code, Cursor, Windsurf & 17+ AI tools.</em></p>
+<p align="center"><code>v3.3.6</code> — Install once. Every session remembers the last. Automatically.</p>
 
 <p align="center">
   <code>+16pp vs Mem0 (zero cloud)</code> &nbsp;·&nbsp; <code>85% Open-Domain (best of any system)</code> &nbsp;·&nbsp; <code>EU AI Act Ready</code>

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.3.6",
+  "version": "3.3.8",
   "description": "Information-geometric agent memory with mathematical guarantees. 4-channel retrieval, Fisher-Rao similarity, zero-LLM mode, EU AI Act compliant. Works with Claude, Cursor, Windsurf, and 17+ AI tools.",
   "keywords": [
     "ai-memory",

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.3.6"
+version = "3.3.8"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "MIT"}

package/src/superlocalmemory/core/config.py

@@ -310,6 +310,7 @@ class PolarQuantConfig:
     dimension: int = 768
     rotation_matrix_path: str = ""  # empty = ~/.superlocalmemory/polar_rotation.npy
     seed: int = 42                  # reproducible rotation matrix
+    codebook_method: str = "turbo"  # "turbo" (default) or "polar_legacy"
 
 
 @dataclass(frozen=True)
@@ -338,7 +339,7 @@ class QuantizationConfig:
     eap_enabled: bool = True
     keep_float32_backup: bool = True
     auto_compact_interval_hours: int = 6
-    polar_search_penalty: float = 0.95
+    polar_search_penalty: float = 0.97  # V3.3.8: 0.95→0.97, TurboQuant has lower MSE
 
 
 @dataclass(frozen=True)

package/src/superlocalmemory/core/embedding_worker.py

@@ -23,9 +23,10 @@ Part of Qualixar | Author: Varun Pratap Bhardwaj
 from __future__ import annotations
 
 import json
+import os
 import signal
 import sys
-import os
+import threading
 
 # Force CPU BEFORE any torch import
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
@@ -41,8 +42,33 @@ if sys.platform != "win32":
     signal.signal(signal.SIGTERM, lambda *_: sys.exit(0))
 
 
+def _start_parent_watchdog() -> None:
+    """Monitor parent process — self-terminate if parent dies.
+
+    Prevents orphaned workers that consume 500-800 MB each when the parent
+    process crashes, is killed, or exits without cleanup.
+
+    V3.3.7: Added after incident where orphaned workers consumed 33 GB.
+    """
+    parent_pid = os.getppid()
+
+    def _watch() -> None:
+        import time
+        while True:
+            time.sleep(5)
+            try:
+                os.kill(parent_pid, 0)
+            except OSError:
+                os._exit(0)
+
+    t = threading.Thread(target=_watch, daemon=True, name="parent-watchdog")
+    t.start()
+
+
 def _worker_main() -> None:
     """Main loop: read JSON requests from stdin, write responses to stdout."""
+    _start_parent_watchdog()  # V3.3.7: self-terminate if parent dies
+
     import numpy as np
 
     model = None

package/src/superlocalmemory/core/embeddings.py

@@ -15,6 +15,7 @@ Part of Qualixar | Author: Varun Pratap Bhardwaj
 
 from __future__ import annotations
 
+import atexit
 import json
 import logging
 import os
@@ -22,11 +23,15 @@ import subprocess
 import sys
 import threading
 import time
+import weakref
 from pathlib import Path
 from typing import TYPE_CHECKING
 
 import numpy as np
 
+# Track all live embedding services for atexit cleanup
+_live_embedding_services: set[weakref.ref] = set()
+
 if TYPE_CHECKING:
     from numpy.typing import NDArray
 
@@ -69,6 +74,17 @@ class EmbeddingService:
         self._worker_ready = False
         self._request_count: int = 0
 
+        # Register for atexit cleanup (prevent orphaned workers)
+        ref = weakref.ref(self, _live_embedding_services.discard)
+        _live_embedding_services.add(ref)
+
+    def __del__(self) -> None:
+        """Kill worker subprocess when service is garbage-collected."""
+        try:
+            self._kill_worker()
+        except Exception:
+            pass
+
     @property
     def is_available(self) -> bool:
         """Check if embedding service can produce embeddings."""
@@ -338,3 +354,26 @@ class EmbeddingService:
             raise DimensionMismatchError(
                 f"Embedding dimension {actual} != expected {self._config.dimension}"
             )
+
+
+# ---------------------------------------------------------------------------
+# Module-level atexit: kill ALL embedding workers on process exit
+# ---------------------------------------------------------------------------
+
+def _cleanup_all_embedding_services() -> None:
+    """Kill all embedding worker subprocesses on interpreter exit.
+
+    Prevents orphaned 500-800 MB sentence-transformer workers surviving
+    after parent exits (especially during test runs with parallel agents).
+    """
+    for ref in list(_live_embedding_services):
+        svc = ref()
+        if svc is not None:
+            try:
+                svc._kill_worker()
+            except Exception:
+                pass
+    _live_embedding_services.clear()
+
+
+atexit.register(_cleanup_all_embedding_services)

package/src/superlocalmemory/core/recall_worker.py

@@ -20,6 +20,7 @@ import json
 import os
 import signal
 import sys
+import threading
 
 # Force CPU BEFORE any torch import
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
@@ -34,6 +35,29 @@ os.environ["TORCH_DEVICE"] = "cpu"
 if sys.platform != "win32":
     signal.signal(signal.SIGTERM, lambda *_: sys.exit(0))
 
+
+def _start_parent_watchdog() -> None:
+    """Monitor parent process — self-terminate if parent dies.
+
+    Prevents orphaned workers that consume 500+ MB each when the parent
+    process crashes, is killed, or exits without cleanup.
+
+    V3.3.7: Added after incident where orphaned workers consumed 33 GB.
+    """
+    parent_pid = os.getppid()
+
+    def _watch() -> None:
+        import time
+        while True:
+            time.sleep(5)
+            try:
+                os.kill(parent_pid, 0)
+            except OSError:
+                os._exit(0)
+
+    t = threading.Thread(target=_watch, daemon=True, name="parent-watchdog")
+    t.start()
+
 _engine = None
 
 
@@ -209,6 +233,8 @@ def _handle_status() -> dict:
 
 def _worker_main() -> None:
     """Main loop: read JSON requests from stdin, write responses to stdout."""
+    _start_parent_watchdog()  # V3.3.7: self-terminate if parent dies
+
     for line in sys.stdin:
         line = line.strip()
         if not line:

package/src/superlocalmemory/math/polar_quant.py

@@ -83,13 +83,23 @@ class PolarQuantEncoder:
     HR-09: Angle indices as uint8, packed into bytes.
     """
 
-    __slots__ = ("_config", "_d", "_S", "_codebooks")
+    __slots__ = ("_config", "_d", "_S", "_codebooks", "_turbo", "_use_turbo")
 
     def __init__(self, config: PolarQuantConfig) -> None:
         self._config = config
         self._d = config.dimension
-        self._S = self._load_or_create_rotation_matrix()
-        self._codebooks = self._generate_uniform_codebooks()
+        codebook_method = getattr(config, "codebook_method", "turbo")
+        if codebook_method == "turbo":
+            from superlocalmemory.math.turbo_quant import TurboQuantEncoder
+            self._turbo = TurboQuantEncoder(config)
+            self._S = self._turbo._S
+            self._codebooks = self._generate_uniform_codebooks()  # for legacy decode
+            self._use_turbo = True
+        else:
+            self._turbo = None
+            self._S = self._load_or_create_rotation_matrix()
+            self._codebooks = self._generate_uniform_codebooks()
+            self._use_turbo = False
 
     # -- Rotation matrix (HR-01, HR-02) ------------------------------------
 
@@ -156,14 +166,14 @@ class PolarQuantEncoder:
     # -- Encode ------------------------------------------------------------
 
     def encode(self, embedding: NDArray, bit_width: int = 4) -> QuantizedEmbedding:
-        """Encode a float32 embedding into quantized polar representation.
+        """Encode a float32 embedding into quantized representation.
 
         Args:
             embedding: 1-D float vector of dimension self._d.
             bit_width:  2, 4, or 8.
 
         Returns:
-            QuantizedEmbedding with packed angle indices.
+            QuantizedEmbedding with packed indices.
 
         Raises:
             ValueError: Invalid bit_width or dimension mismatch.
@@ -177,13 +187,25 @@ class PolarQuantEncoder:
                 f"shape mismatch: expected ({self._d},), got {embedding.shape}"
             )
 
-        # Step 1: Random rotation
-        v_rot = self._S @ embedding
+        # V3.3.8: TurboQuant path (default)
+        if self._use_turbo:
+            result = self._turbo.encode(embedding, bit_width)
+            return QuantizedEmbedding(
+                fact_id="",
+                radius=result.radius,
+                angle_indices=result.indices,
+                bit_width=result.bit_width,
+                qjl_bits=None,
+            )
 
-        # Step 2: Compute radius
+        # Legacy PolarQuant path
+        return self._encode_polar(embedding, bit_width)
+
+    def _encode_polar(self, embedding: NDArray, bit_width: int) -> QuantizedEmbedding:
+        """Legacy PolarQuant encode (polar coordinate transform)."""
+        v_rot = self._S @ embedding
         r = float(np.linalg.norm(v_rot))
 
-        # Degenerate zero vector
         if r < 1e-12:
             zero_angles = np.zeros(self._d - 1, dtype=np.uint8)
             if bit_width == 8:
@@ -200,17 +222,11 @@ class PolarQuantEncoder:
                 qjl_bits=None,
             )
 
-        # Step 3: Normalize
         v_unit = v_rot / r
-
-        # Step 4: Cartesian to polar angles
         angles = _cartesian_to_polar_angles(v_unit)
-
-        # Step 5: Quantize angles using codebook
         cb = self._codebooks[bit_width]
         indices = np.digitize(angles, cb["boundaries"][1:-1]).astype(np.uint8)
 
-        # Step 6: Pack into bytes
         if bit_width == 8:
             packed = indices.tobytes()
         elif bit_width == 4:
@@ -228,18 +244,43 @@ class PolarQuantEncoder:
 
     # -- Decode ------------------------------------------------------------
 
+    # TQ magic prefix for format detection (HR-MIG-02)
+    _TQ_MAGIC = b"\x54\x51"
+
     def decode(self, qe: QuantizedEmbedding) -> NDArray:
         """Decode a QuantizedEmbedding back to float64 vector.
 
+        V3.3.8: Detects "TQ" prefix (0x54, 0x51) to route between
+        TurboQuant and legacy PolarQuant decode paths.
+
         Args:
             qe: Quantized embedding produced by encode().
 
         Returns:
             Reconstructed vector of dimension self._d.
         """
+        # Format detection: TQ prefix = TurboQuant, else legacy polar
+        if qe.angle_indices[:2] == self._TQ_MAGIC:
+            return self._decode_turbo(qe)
+        return self._decode_polar(qe)
+
+    def _decode_turbo(self, qe: QuantizedEmbedding) -> NDArray:
+        """Decode TurboQuant-encoded BLOB (has TQ prefix)."""
+        if self._turbo is None:
+            from superlocalmemory.math.turbo_quant import TurboQuantEncoder
+            self._turbo = TurboQuantEncoder(self._config)
+        from superlocalmemory.math.turbo_quant import TurboQuantResult
+        result = TurboQuantResult(
+            radius=qe.radius,
+            indices=qe.angle_indices,
+            bit_width=qe.bit_width,
+        )
+        return self._turbo.decode(result)
+
+    def _decode_polar(self, qe: QuantizedEmbedding) -> NDArray:
+        """Decode legacy PolarQuant BLOB (no TQ prefix)."""
         n_angles = self._d - 1
 
-        # Step 1: Unpack angle indices
         if qe.bit_width == 8:
             indices = np.frombuffer(qe.angle_indices, dtype=np.uint8).copy()
         elif qe.bit_width == 4:
@@ -247,19 +288,12 @@ class PolarQuantEncoder:
         else:
             indices = self.unpack_2bit(qe.angle_indices, n_angles)
 
-        # Step 2: Dequantize -- map indices to centroid angles
         centroids = self._codebooks[qe.bit_width]["centroids"]
-        # Clip indices to valid range
         indices = np.clip(indices, 0, len(centroids) - 1)
         angles = centroids[indices]
 
-        # Step 3: Polar to Cartesian
         v_unit = _polar_to_cartesian(angles, self._d)
-
-        # Step 4: Scale by radius
         v_rot = v_unit * qe.radius
-
-        # Step 5: Inverse rotation (S is orthogonal, so S^T = S^{-1})
         v_orig = self._S.T @ v_rot
 
         return v_orig

package/src/superlocalmemory/math/turbo_quant.py

@@ -0,0 +1,308 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under the MIT License - see LICENSE file
+# Part of SuperLocalMemory V3
+
+"""TurboQuant embedding quantization (ICLR 2026).
+
+Per-coordinate Lloyd-Max scalar quantization after random orthogonal rotation.
+D_mse <= sqrt(3*pi/2) / 4^b. No scipy (HR-SCIPY-01). 2-byte "TQ" prefix on
+all BLOBs (HR-MIG-02). Bit-widths: 2, 4, 8 only (HR-3BIT-01).
+
+References: TurboQuant (arXiv 2504.19874), PolarQuant (arXiv 2502.02617).
+Part of Qualixar | Author: Varun Pratap Bhardwaj | License: MIT
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import shutil
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+from numpy.typing import NDArray
+
+from superlocalmemory.core.config import PolarQuantConfig
+
+logger = logging.getLogger(__name__)
+
+TQ_MAGIC = b"\x54\x51"  # 2-byte prefix for TurboQuant BLOBs (HR-MIG-02)
+SUPPORTED_BIT_WIDTHS: frozenset[int] = frozenset({2, 4, 8})
+
+# ---------------------------------------------------------------------------
+# Data types
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class TurboQuantResult:
+    """Immutable TurboQuant-encoded embedding. radius=float16, indices=TQ-prefixed."""
+
+    radius: float
+    indices: bytes  # TQ_MAGIC + packed codebook indices
+    bit_width: int
+
+
+# ---------------------------------------------------------------------------
+# Lloyd-Max codebook (HR-SCIPY-01: math.erf + math.exp only)
+# ---------------------------------------------------------------------------
+
+_SQRT_2PI = math.sqrt(2.0 * math.pi)
+_SQRT_2 = math.sqrt(2.0)
+
+
+def _std_normal_pdf(x: float) -> float:
+    return math.exp(-0.5 * x * x) / _SQRT_2PI
+
+
+def _std_normal_cdf(x: float) -> float:
+    return 0.5 * (1.0 + math.erf(x / _SQRT_2))
+
+
+def _compute_lloyd_max_gaussian(
+    sigma: float, n_levels: int, max_iter: int = 100, tol: float = 1e-10,
+) -> NDArray:
+    """Lloyd-Max optimal codebook for N(0, sigma^2). Deterministic (HR-CB-01)."""
+    lo, hi = -5.0 * sigma, 5.0 * sigma
+    boundaries = np.linspace(lo, hi, n_levels + 1)
+    centroids = np.zeros(n_levels)
+    for k in range(n_levels):
+        centroids[k] = 0.5 * (boundaries[k] + boundaries[k + 1])
+
+    for _ in range(max_iter):
+        old = centroids.copy()
+        for k in range(n_levels):
+            a_k = float(boundaries[k]) / sigma
+            b_k = float(boundaries[k + 1]) / sigma
+            denom = _std_normal_cdf(b_k) - _std_normal_cdf(a_k)
+            if denom > 1e-15:
+                centroids[k] = sigma * (_std_normal_pdf(a_k) - _std_normal_pdf(b_k)) / denom
+        for k in range(1, n_levels):
+            boundaries[k] = 0.5 * (centroids[k - 1] + centroids[k])
+        if float(np.max(np.abs(centroids - old))) < tol:
+            break
+
+    return np.sort(centroids)
+
+
+# ---------------------------------------------------------------------------
+# Bit packing
+# ---------------------------------------------------------------------------
+
+
+def _pack_8bit(indices: NDArray) -> bytes:
+    return indices.astype(np.uint8).tobytes()
+
+
+def _unpack_8bit(data: bytes, length: int) -> NDArray:
+    return np.frombuffer(data, dtype=np.uint8)[:length].copy()
+
+
+def _pack_4bit(indices: NDArray) -> bytes:
+    n = len(indices)
+    padded = np.zeros(n + (n % 2), dtype=np.uint8)
+    padded[:n] = np.clip(indices, 0, 15)
+    return ((padded[0::2] << 4) | padded[1::2]).tobytes()
+
+
+def _unpack_4bit(data: bytes, length: int) -> NDArray:
+    packed = np.frombuffer(data, dtype=np.uint8)
+    result = np.empty(len(packed) * 2, dtype=np.uint8)
+    result[0::2] = packed >> 4
+    result[1::2] = packed & 0x0F
+    return result[:length]
+
+
+def _pack_2bit(indices: NDArray) -> bytes:
+    n = len(indices)
+    padded = np.zeros(n + (4 - n % 4) % 4, dtype=np.uint8)
+    padded[:n] = np.clip(indices, 0, 3)
+    return (
+        (padded[0::4] << 6) | (padded[1::4] << 4)
+        | (padded[2::4] << 2) | padded[3::4]
+    ).tobytes()
+
+
+def _unpack_2bit(data: bytes, length: int) -> NDArray:
+    packed = np.frombuffer(data, dtype=np.uint8)
+    result = np.empty(len(packed) * 4, dtype=np.uint8)
+    result[0::4] = (packed >> 6) & 0x03
+    result[1::4] = (packed >> 4) & 0x03
+    result[2::4] = (packed >> 2) & 0x03
+    result[3::4] = packed & 0x03
+    return result[:length]
+
+
+_PACKERS: dict[int, tuple] = {
+    8: (_pack_8bit, _unpack_8bit),
+    4: (_pack_4bit, _unpack_4bit),
+    2: (_pack_2bit, _unpack_2bit),
+}
+
+# ---------------------------------------------------------------------------
+# TurboQuantEncoder
+# ---------------------------------------------------------------------------
+
+
+class TurboQuantEncoder:
+    """Per-coordinate Lloyd-Max quantizer with random rotation.
+
+    HR-ROT-01: Same rotation matrix for encode/decode.
+    HR-CB-02: Codebooks computed ONCE at __init__.
+    HR-SCIPY-01: No scipy dependency.
+    """
+
+    __slots__ = ("_config", "_d", "_S", "_codebooks")
+
+    def __init__(self, config: PolarQuantConfig) -> None:
+        self._config = config
+        self._d = config.dimension
+        self._S = self._load_or_create_rotation_matrix()
+        self._codebooks = self._compute_codebooks()
+
+    def _load_or_create_rotation_matrix(self) -> NDArray:
+        """Load/create rotation matrix with copy-on-detect (AUDIT C4-MED-01)."""
+        d = self._d
+        slm_dir = Path.home() / ".superlocalmemory"
+
+        turbo_path_str = self._config.rotation_matrix_path
+        if not turbo_path_str:
+            turbo_path_str = str(slm_dir / f"turbo_rotation_{d}.npy")
+        turbo_path = Path(turbo_path_str)
+
+        if turbo_path.exists():
+            try:
+                S = np.load(str(turbo_path))
+                if S.shape == (d, d):
+                    return S
+                logger.warning("Turbo rotation shape %s != (%d,%d)", S.shape, d, d)
+            except Exception as exc:
+                logger.warning("Corrupt turbo rotation: %s", exc)
+
+        # Copy-on-detect: reuse existing polar rotation matrix
+        polar_path = slm_dir / f"polar_rotation_{d}.npy"
+        if polar_path.exists() and not turbo_path.exists():
+            try:
+                S = np.load(str(polar_path))
+                if S.shape == (d, d):
+                    turbo_path.parent.mkdir(parents=True, exist_ok=True)
+                    shutil.copy2(str(polar_path), str(turbo_path))
+                    logger.info("Copied polar rotation matrix for TurboQuant compatibility")
+                    return S
+            except Exception as exc:
+                logger.warning("Could not copy polar rotation: %s", exc)
+
+        # Generate new via Mezzadri-corrected QR
+        rng = np.random.default_rng(self._config.seed)
+        H = rng.standard_normal((d, d))
+        Q, R = np.linalg.qr(H)
+        S = Q @ np.diag(np.sign(np.diag(R)))
+
+        turbo_path.parent.mkdir(parents=True, exist_ok=True)
+        np.save(str(turbo_path), S)
+        logger.info("Generated TurboQuant rotation (%d x %d) at %s", d, d, turbo_path)
+        return S
+
+    def _compute_codebooks(self) -> dict[int, NDArray]:
+        """Pre-compute Lloyd-Max codebooks for 2/4/8-bit."""
+        sigma = 1.0 / math.sqrt(self._d)
+        codebooks: dict[int, NDArray] = {}
+        for bw in sorted(SUPPORTED_BIT_WIDTHS):
+            centroids = _compute_lloyd_max_gaussian(sigma, 2 ** bw)
+            assert len(centroids) == 2 ** bw
+            assert np.all(centroids[1:] >= centroids[:-1])
+            codebooks[bw] = centroids
+        return codebooks
+
+    def encode(self, embedding: NDArray, bit_width: int = 4) -> TurboQuantResult:
+        """Encode embedding. HR-ENC-01: pure. HR-ENC-02: radius=float16."""
+        if bit_width not in SUPPORTED_BIT_WIDTHS:
+            raise ValueError(f"bit_width must be 2, 4, or 8, got {bit_width}")
+        if embedding.shape != (self._d,):
+            raise ValueError(f"shape mismatch: expected ({self._d},), got {embedding.shape}")
+
+        y = self._S @ embedding
+        r = float(np.linalg.norm(y))
+
+        if r < 1e-12:
+            pack_fn, _ = _PACKERS[bit_width]
+            packed = TQ_MAGIC + pack_fn(np.zeros(self._d, dtype=np.uint8))
+            return TurboQuantResult(radius=0.0, indices=packed, bit_width=bit_width)
+
+        y_unit = y / r
+        centroids = self._codebooks[bit_width]
+        idx = np.searchsorted(centroids, y_unit)
+        idx = np.clip(idx, 0, len(centroids) - 1)
+        left = np.clip(idx - 1, 0, len(centroids) - 1)
+        use_left = np.abs(y_unit - centroids[left]) < np.abs(y_unit - centroids[idx])
+        idx = np.where(use_left, left, idx).astype(np.uint8)
+
+        pack_fn, _ = _PACKERS[bit_width]
+        packed = TQ_MAGIC + pack_fn(idx)
+
+        return TurboQuantResult(
+            radius=float(np.float16(r)), indices=packed, bit_width=bit_width,
+        )
+
+    def decode(self, result: TurboQuantResult) -> NDArray:
+        """Decode with format detection: TQ prefix -> turbo, else -> legacy polar."""
+        blob = result.indices
+
+        if blob[:2] == TQ_MAGIC:
+            data = blob[2:]
+        else:
+            return self._decode_legacy_polar(result)
+
+        _, unpack_fn = _PACKERS[result.bit_width]
+        indices = unpack_fn(data, self._d)
+        centroids = self._codebooks[result.bit_width]
+        y_unit_approx = centroids[np.clip(indices, 0, len(centroids) - 1)]
+        return self._S.T @ (y_unit_approx * result.radius)
+
+    def _decode_legacy_polar(self, result: TurboQuantResult) -> NDArray:
+        """Decode legacy PolarQuant BLOB (no TQ prefix) for SLM <= 3.3.6."""
+        from superlocalmemory.math.polar_quant import PolarQuantEncoder, _polar_to_cartesian
+
+        n_angles = self._d - 1
+        if result.bit_width == 8:
+            indices = np.frombuffer(result.indices, dtype=np.uint8).copy()
+        elif result.bit_width == 4:
+            indices = PolarQuantEncoder.unpack_4bit(result.indices, n_angles)
+        else:
+            indices = PolarQuantEncoder.unpack_2bit(result.indices, n_angles)
+
+        levels = 2 ** result.bit_width
+        boundaries = np.linspace(0.0, math.pi, levels + 1)
+        centroids = (boundaries[:-1] + boundaries[1:]) / 2.0
+        angles = centroids[np.clip(indices, 0, len(centroids) - 1)]
+
+        v_unit = _polar_to_cartesian(angles, self._d)
+        return self._S.T @ (v_unit * result.radius)
+
+    def approximate_similarity(self, query: NDArray, result: TurboQuantResult) -> float:
+        """Cosine similarity via decode. Returns 0.0 on degenerate inputs."""
+        decoded = self.decode(result)
+        denom = np.linalg.norm(query) * np.linalg.norm(decoded)
+        if denom < 1e-12:
+            return 0.0
+        sim = float(np.dot(query, decoded) / denom)
+        return 0.0 if (math.isnan(sim) or math.isinf(sim)) else sim
+
+    # Static pack/unpack (backward compat with PolarQuantEncoder API)
+
+    @staticmethod
+    def pack_4bit(indices: NDArray) -> bytes:
+        return _pack_4bit(indices)
+
+    @staticmethod
+    def unpack_4bit(data: bytes, length: int) -> NDArray:
+        return _unpack_4bit(data, length)
+
+    @staticmethod
+    def pack_2bit(indices: NDArray) -> bytes:
+        return _pack_2bit(indices)
+
+    @staticmethod
+    def unpack_2bit(data: bytes, length: int) -> NDArray:
+        return _unpack_2bit(data, length)

package/src/superlocalmemory/retrieval/reranker.py

@@ -16,6 +16,7 @@ License: MIT
 
 from __future__ import annotations
 
+import atexit
 import json
 import logging
 import os
@@ -23,10 +24,14 @@ import subprocess
 import sys
 import threading
 import time
+import weakref
 from typing import Any
 
 from superlocalmemory.storage.models import AtomicFact
 
+# Track all live reranker instances for atexit cleanup
+_live_rerankers: set[weakref.ref] = set()
+
 logger = logging.getLogger(__name__)
 
 _IDLE_TIMEOUT_SECONDS = 120  # 2 min → kill worker
@@ -64,11 +69,22 @@ class CrossEncoderReranker:
         self._idle_timer: threading.Timer | None = None
         self._request_count: int = 0
 
+        # Register for atexit cleanup (prevent orphaned workers)
+        ref = weakref.ref(self, _live_rerankers.discard)
+        _live_rerankers.add(ref)
+
         # Start background warmup immediately — worker loads model
         # while the rest of init continues. First recall gets instant
         # fallback; second recall uses the warm model.
         self._start_background_warmup()
 
+    def __del__(self) -> None:
+        """Kill worker subprocess when reranker is garbage-collected."""
+        try:
+            self._kill_worker()
+        except Exception:
+            pass
+
     # ------------------------------------------------------------------
     # Background warmup (non-blocking model load)
     # ------------------------------------------------------------------
@@ -330,3 +346,26 @@ class CrossEncoderReranker:
         if resp is None:
             return False
         return resp.get("ok", False)
+
+
+# ---------------------------------------------------------------------------
+# Module-level atexit: kill ALL reranker workers on process exit
+# ---------------------------------------------------------------------------
+
+def _cleanup_all_rerankers() -> None:
+    """Kill all reranker worker subprocesses on interpreter exit.
+
+    Prevents orphaned 1.3 GB ONNX/PyTorch workers surviving after
+    parent exits (especially during test runs with parallel agents).
+    """
+    for ref in list(_live_rerankers):
+        reranker = ref()
+        if reranker is not None:
+            try:
+                reranker._kill_worker()
+            except Exception:
+                pass
+    _live_rerankers.clear()
+
+
+atexit.register(_cleanup_all_rerankers)