add AGX_RELAX_CDM_CTXSTORE_TIMEOUT driver workaround

Relaxes Metal command buffer context store timeout to reduce kernel
panics on long-running GPU workloads. Zero-cost env var hint to the
IOGPUFamily driver, set at import time.

Suggested by @zcbenz (MLX maintainer) in ml-explore/mlx#3267.
Surfaced by Harperbot/metal-guard.

Bump to v1.0.1.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: alyssapowell

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "mlx-halo"
-version = "1.0.0"
+version = "1.0.1"
 description = "Pre-flight safety checks for MLX models on Apple Silicon. Prevents kernel panics from overlapping Metal GPU allocations."
 license = {file = "LICENSE.md"}
 requires-python = ">=3.10"

src/mlx_halo/__init__.py

@@ -23,13 +23,25 @@
     result = halo.check_all(estimated_model_gb=18.0)
 """
 
+import os as _os
+
+# ── Apple GPU driver workaround ──────────────────────────────────────────
+# Relaxes the Metal command buffer context store timeout to reduce kernel
+# panics on long-running GPU workloads. Zero-cost env var hint to the
+# IOGPUFamily driver — safe to set unconditionally.
+#
+# Suggested by @zcbenz (MLX maintainer) in ml-explore/mlx#3267.
+# Surfaced by Harperbot/metal-guard (runtime MLX safety layer).
+if "AGX_RELAX_CDM_CTXSTORE_TIMEOUT" not in _os.environ:
+    _os.environ["AGX_RELAX_CDM_CTXSTORE_TIMEOUT"] = "1"
+
 from .safety import HaloCheck, preflight
 from .monitor import SystemMonitor, get_monitor
 from .pain import PainCalculator, get_pain_calculator, get_current_pain
 from .memory import get_gpu_memory_status, clear_gpu_cache, wait_for_memory_drain
 from .types import SystemMetrics, PainProfile, HaloResult, HealthStatus, MemoryStatus
 
-__version__ = "1.0.0"
+__version__ = "1.0.1"
 
 __all__ = [
     "preflight",