Make zarr/numcodecs imports optional for ARM64 compatibility

- Use try/except for zarr imports
- Add ZARR_AVAILABLE flag
- Skip storage operations gracefully when zarr not available
- Core model inference still works without tensor caching

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

backend/storage.py

@@ -16,8 +16,16 @@ Storage structure:
 └── telemetry.jsonl
 """
 
-import zarr
-import numcodecs
+# Optional imports for zarr storage (not available on ARM64)
+try:
+    import zarr
+    import numcodecs
+    ZARR_AVAILABLE = True
+except ImportError:
+    zarr = None
+    numcodecs = None
+    ZARR_AVAILABLE = False
+
 import numpy as np
 import torch
 import json
@@ -53,12 +61,16 @@ class ZarrStorage:
         # Create directories
         self.tensor_dir.mkdir(parents=True, exist_ok=True)
 
-        # Blosc compressor for efficient compression
-        self.compressor = numcodecs.Blosc(
-            cname='zstd',      # zstd algorithm (good compression + speed)
-            clevel=5,          # Compression level (1-9, 5 is balanced)
-            shuffle=numcodecs.Blosc.SHUFFLE  # Byte shuffle for better compression
-        )
+        # Blosc compressor for efficient compression (if zarr available)
+        if ZARR_AVAILABLE:
+            self.compressor = numcodecs.Blosc(
+                cname='zstd',      # zstd algorithm (good compression + speed)
+                clevel=5,          # Compression level (1-9, 5 is balanced)
+                shuffle=numcodecs.Blosc.SHUFFLE  # Byte shuffle for better compression
+            )
+        else:
+            self.compressor = None
+            logger.warning("Zarr not available - tensor storage will be disabled")
 
     def save_instrumentation_data(self, data: InstrumentationData) -> Dict[str, Any]:
         """
@@ -70,6 +82,10 @@ class ZarrStorage:
         Returns:
             Dictionary with file paths and sizes
         """
+        if not ZARR_AVAILABLE:
+            logger.warning("Zarr not available - skipping instrumentation data save")
+            return {'run_id': self.run_id, 'error': 'zarr_not_available'}
+
         logger.info(f"Saving instrumentation data for run {self.run_id}...")
 
         result = {
@@ -227,6 +243,9 @@ class ZarrStorage:
         Returns:
             Attention matrix [seq_len, seq_len]
         """
+        if not ZARR_AVAILABLE:
+            raise RuntimeError("Zarr not available - cannot load attention data")
+
         path = self.tensor_dir / "attention.zarr"
 
         if not path.exists():