Fix session persistence: pickle VAD artifacts to preserve tensor types

clean_speech_intervals() expects the original torch.Tensor types from
VAD output. The previous np.save/np.load roundtrip converted them to
numpy arrays, causing resegment_session to crash. Now uses pickle for
speech_intervals and is_complete, keeping np.save only for audio.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/api/session_api.py

@@ -8,6 +8,7 @@ re-uploads and re-inference.
 import hashlib
 import json
 import os
+import pickle
 import re
 import shutil
 import time
@@ -36,24 +37,8 @@ def _validate_id(audio_id: str) -> bool:
     return isinstance(audio_id, str) and bool(_VALID_ID.match(audio_id))
 
 
-def _is_expired(meta: dict) -> bool:
-    return (time.time() - meta.get("created_at", 0)) > SESSION_EXPIRY_SECONDS
-
-
-def _read_metadata(session_path):
-    meta_path = session_path / "metadata.json"
-    if not meta_path.exists():
-        return None
-    with open(meta_path) as f:
-        return json.load(f)
-
-
-def _write_metadata(session_path, meta: dict):
-    """Atomic write via temp file + os.replace."""
-    tmp = session_path / "metadata.tmp"
-    with open(tmp, "w") as f:
-        json.dump(meta, f)
-    os.replace(tmp, session_path / "metadata.json")
+def _is_expired(created_at: float) -> bool:
+    return (time.time() - created_at) > SESSION_EXPIRY_SECONDS
 
 
 def _sweep_expired():
@@ -68,8 +53,8 @@ def _sweep_expired():
     for entry in SESSION_DIR.iterdir():
         if not entry.is_dir():
             continue
-        meta = _read_metadata(entry)
-        if meta is None or _is_expired(meta):
+        ts_file = entry / "created_at"
+        if not ts_file.exists() or _is_expired(float(ts_file.read_text())):
             shutil.rmtree(entry, ignore_errors=True)
 
 
@@ -78,23 +63,37 @@ def _intervals_hash(intervals) -> str:
 
 
 def create_session(audio, speech_intervals, is_complete, intervals, model_name):
-    """Persist session data and return audio_id (32-char hex UUID)."""
+    """Persist session data and return audio_id (32-char hex UUID).
+
+    Uses pickle for VAD artifacts (speech_intervals, is_complete) to
+    preserve exact types (torch.Tensor etc.) expected by the segmenter.
+    Uses np.save for the audio array (large, always float32 numpy).
+    """
     _sweep_expired()
     audio_id = uuid.uuid4().hex
     path = _session_dir(audio_id)
     path.mkdir(parents=True, exist_ok=True)
 
+    # Audio is always a float32 numpy array after preprocessing
     np.save(path / "audio.npy", audio)
-    np.save(path / "speech_intervals.npy", speech_intervals)
 
+    # VAD artifacts: preserve original types via pickle
+    with open(path / "vad.pkl", "wb") as f:
+        pickle.dump({"speech_intervals": speech_intervals,
+                      "is_complete": is_complete}, f)
+
+    # Lightweight metadata (JSON-safe types only)
     meta = {
-        "is_complete": bool(is_complete),
         "intervals": intervals,
         "model_name": model_name,
         "intervals_hash": _intervals_hash(intervals),
-        "created_at": time.time(),
     }
-    _write_metadata(path, meta)
+    with open(path / "metadata.json", "w") as f:
+        json.dump(meta, f)
+
+    # Timestamp file for cheap expiry checks during sweep
+    (path / "created_at").write_text(str(time.time()))
+
     return audio_id
 
 
@@ -105,18 +104,24 @@ def load_session(audio_id):
     path = _session_dir(audio_id)
     if not path.exists():
         return None
-    meta = _read_metadata(path)
-    if meta is None or _is_expired(meta):
+
+    ts_file = path / "created_at"
+    if not ts_file.exists() or _is_expired(float(ts_file.read_text())):
         shutil.rmtree(path, ignore_errors=True)
         return None
 
     audio = np.load(path / "audio.npy")
-    speech_intervals = np.load(path / "speech_intervals.npy")
+
+    with open(path / "vad.pkl", "rb") as f:
+        vad = pickle.load(f)
+
+    with open(path / "metadata.json") as f:
+        meta = json.load(f)
 
     return {
         "audio": audio,
-        "speech_intervals": speech_intervals,
-        "is_complete": meta["is_complete"],
+        "speech_intervals": vad["speech_intervals"],
+        "is_complete": vad["is_complete"],
         "intervals": meta["intervals"],
         "model_name": meta["model_name"],
         "intervals_hash": meta.get("intervals_hash", ""),
@@ -127,15 +132,20 @@ def load_session(audio_id):
 def update_session(audio_id, *, intervals=None, model_name=None):
     """Update mutable session fields (intervals, model_name)."""
     path = _session_dir(audio_id)
-    meta = _read_metadata(path)
-    if meta is None:
+    meta_path = path / "metadata.json"
+    if not meta_path.exists():
         return
+    with open(meta_path) as f:
+        meta = json.load(f)
     if intervals is not None:
         meta["intervals"] = intervals
         meta["intervals_hash"] = _intervals_hash(intervals)
     if model_name is not None:
         meta["model_name"] = model_name
-    _write_metadata(path, meta)
+    tmp = path / "metadata.tmp"
+    with open(tmp, "w") as f:
+        json.dump(meta, f)
+    os.replace(tmp, meta_path)
 
 
 # ---------------------------------------------------------------------------