Add native Windows compatibility layer

src/embeddings.py

@@ -13,6 +13,17 @@ Set EMBEDDING_URL in .env, e.g.:
 """
 
 import os
+
+# Windows: force HuggingFace/fastembed to COPY model files rather than symlink
+# them. On a network-share/UNC cache dir Windows can't follow HF's symlinks
+# ([WinError 1463] "symbolic link cannot be followed"), so ONNX fails to load the
+# model and semantic memory dies. huggingface_hub reads this flag at import time,
+# so it must be set before huggingface_hub is first imported — hence module-top.
+# (app.py sets the same guard for the server entrypoint.)
+if os.name == "nt":
+    os.environ.setdefault("HF_HUB_DISABLE_SYMLINKS", "1")
+    os.environ.setdefault("HF_HUB_DISABLE_SYMLINKS_WARNING", "1")
+
 import logging
 import numpy as np
 import httpx
@@ -109,6 +120,35 @@ class FastEmbedClient:
             "data", "fastembed_cache",
         )
         os.makedirs(cache_dir, exist_ok=True)
+        # Windows self-heal: the HuggingFace-hub cache stores model files as
+        # symlinks (snapshots/<rev>/model.onnx -> ../../blobs/<hash>). On a
+        # network-share / UNC data dir Windows refuses to follow them
+        # ([WinError 1463] "symbolic link cannot be followed because its type is
+        # disabled"), and a cache copied between machines can carry dead symlinks
+        # too. Either way fastembed tries to load a broken symlink and fails
+        # *without* re-downloading, leaving semantic memory degraded. Detect a
+        # broken-symlink model in the cache and drop the contaminated hub dir so
+        # fastembed re-fetches (it falls back to its CDN tarball of real files,
+        # which load fine). Best-effort; only ever removes a verifiably dead link.
+        if os.name == "nt":
+            try:
+                import glob, shutil
+                for _onnx in glob.glob(os.path.join(cache_dir, "**", "*.onnx"), recursive=True):
+                    if os.path.islink(_onnx) and not os.path.exists(_onnx):
+                        _root = _onnx
+                        while os.path.basename(_root) and not os.path.basename(_root).startswith("models--"):
+                            _parent = os.path.dirname(_root)
+                            if _parent == _root:
+                                break
+                            _root = _parent
+                        if os.path.basename(_root).startswith("models--"):
+                            logger.warning(
+                                "Embedding cache has a broken symlink (%s); clearing %s "
+                                "so fastembed re-downloads real files", _onnx, _root,
+                            )
+                            shutil.rmtree(_root, ignore_errors=True)
+            except Exception as _e:
+                logger.debug("embedding cache symlink-heal skipped: %s", _e)
         kwargs = {"model_name": self.model, "cache_dir": cache_dir}
         self._embedding = TextEmbedding(**kwargs)
         self._dim: Optional[int] = None
@@ -152,7 +192,7 @@ def _load_persisted_endpoint() -> dict:
         )
         if os.path.exists(endpoint_file):
             import json
-            data = json.loads(open(endpoint_file).read())
+            data = json.loads(open(endpoint_file, encoding="utf-8").read())
             if data.get("url"):
                 return data
     except Exception: