superlocalmemory 3.4.59 → 3.4.60

package/CHANGELOG.md

@@ -5,6 +5,39 @@ All notable changes to SuperLocalMemory V3 will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.4.60] - 2026-05-31 — Daemon OpenMP Crash Hotfix
+
+**Hotfix for v3.4.59.** Forces `OMP_NUM_THREADS=1` and `KMP_DUPLICATE_LIB_OK=TRUE`
+in the daemon subprocess environment BEFORE Python imports any C extensions that
+bundle their own libomp.dylib (torch, scikit-learn, lightgbm).
+
+### Root Cause
+Setting these env vars in `superlocalmemory/__init__.py` (v3.4.58 fix) was too
+late on Apple Silicon (M5 Pro). By the time `import superlocalmemory` runs, the
+parent process has often already loaded one of the OpenMP-using extensions, and
+that libomp's thread pool is initialized. When lightgbm later forks its worker
+pool from `LGBM_DatasetCreateFromMat`, the parent's libomp thread structs are
+incompatible with lightgbm's libomp → SIGSEGV at
+`__kmp_suspend_initialize_thread` reading address `0x580`.
+
+### Fix
+`cli/daemon.py:_start_daemon_subprocess()` now copies `os.environ`, sets
+`OMP_NUM_THREADS=1` + `KMP_DUPLICATE_LIB_OK=TRUE`, and passes it via the
+`env=` kwarg to `subprocess.Popen`. The daemon Python interpreter starts
+with these vars already in its environment, so C extension imports see the
+correct values during their `_init` constructors.
+
+### Why 1 thread and not 2
+v3.4.58 set `OMP_NUM_THREADS=2` as a compromise. On M5 Pro the parallel-fork
+race still triggered at 2. SLM's actual ML workloads (ranker retrain, dedup)
+operate on datasets of 50–5,000 rows where 1-thread serial OpenMP is within
+~10% of multi-threaded but eliminates the crash class entirely.
+
+### Changed
+- `cli/daemon.py`: `_start_daemon_subprocess()` injects `env={OMP_NUM_THREADS=1, KMP_DUPLICATE_LIB_OK=TRUE}` into Popen
+
+---
+
 ## [3.4.59] - 2026-05-31 — Graph Edge Cap + Recall Reliability
 
 **Fixes SLM falling into degraded FTS5 mode on every session start**, and stops

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.4.59",
+  "version": "3.4.60",
   "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",
@@ -88,4 +88,4 @@
   "dependencies": {
     "docx": "^9.5.1"
   }
-}
+}

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.4.59"
+version = "3.4.60"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "AGPL-3.0-or-later"}

package/src/superlocalmemory/__init__.py

@@ -28,7 +28,7 @@ if "OMP_NUM_THREADS" not in os.environ:
     os.environ["OMP_NUM_THREADS"] = "2"
 # ---------------------------------------------------------------------------
 
-__version__ = "3.4.59"
+__version__ = "3.4.60"
 
 _REQUIRED_VERSIONS = {
     "sentence_transformers": "5.3.0",

package/src/superlocalmemory/cli/daemon.py

@@ -171,6 +171,17 @@ def _start_daemon_subprocess() -> bool:
     else:
         kwargs["start_new_session"] = True
 
+    # v3.4.60: Force OMP_NUM_THREADS=1 in daemon env BEFORE Python imports
+    # numpy/torch/lightgbm. Setting it in __init__.py is too late on M5 Pro —
+    # by the time superlocalmemory.__init__ runs, libomp has already been
+    # initialized by an earlier import, causing the SIGSEGV at
+    # __kmp_suspend_initialize_thread when lightgbm forks its worker pool.
+    # Forcing serial OpenMP eliminates the parallel barrier race entirely.
+    daemon_env = os.environ.copy()
+    daemon_env["OMP_NUM_THREADS"] = "1"
+    daemon_env["KMP_DUPLICATE_LIB_OK"] = "TRUE"
+    kwargs["env"] = daemon_env
+
     with open(log_file, "a") as lf:
         proc = subprocess.Popen(cmd, stdout=lf, stderr=lf, **kwargs)
 

package/src/superlocalmemory.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 3.4.59
+Version: 3.4.60
 Summary: Information-geometric agent memory with mathematical guarantees
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License: AGPL-3.0-or-later