feat: add thread lock to prevent concurrent model corruption

Add threading lock around model.generate() to prevent concurrent calls from corrupting shared model state (alignment_stream_analyzer). Update comments to explain the concurrency issue and upgrade path.

app.py

@@ -1,6 +1,7 @@
 import os
 import re
 import secrets
+import threading
 import uuid
 
 from fastapi import Depends, FastAPI, HTTPException, Header, Request
@@ -14,6 +15,11 @@ os.makedirs(OUT_DIR, exist_ok=True)
 
 app = FastAPI(title="Chatterbox TTS")
 _model = None
+# ponytail: global lock. chatterbox keeps per-generation state on the shared model
+# (alignment_stream_analyzer), so concurrent generate() calls corrupt each other.
+# Sync endpoints run in a threadpool, so "one worker" is not one caller.
+# Upgrade path: a real request queue / model pool if throughput matters.
+_gen_lock = threading.Lock()
 
 
 def auth(x_api_key: str = Header(None)):
@@ -23,7 +29,8 @@ def auth(x_api_key: str = Header(None)):
 
 def model():
     # ponytail: lazy singleton so the Space boots fast and healthchecks pass.
-    # Single uvicorn worker, so no lock needed; add one if we ever scale up.
+    # Worst case two requests both build it and one wins; generate() is serialized
+    # by _gen_lock anyway.
     global _model
     if _model is None:
         import functools
@@ -58,12 +65,13 @@ def tts(req: TTSRequest, request: Request):
     if lang not in m.get_supported_languages():
         raise HTTPException(400, f"unsupported lang: {lang}")
 
-    wav = m.generate(
-        req.text,
-        language_id=lang,
-        exaggeration=req.exaggeration,
-        cfg_weight=req.cfg_weight,
-    )
+    with _gen_lock:
+        wav = m.generate(
+            req.text,
+            language_id=lang,
+            exaggeration=req.exaggeration,
+            cfg_weight=req.cfg_weight,
+        )
     stem = uuid.uuid4().hex
     torchaudio.save(os.path.join(OUT_DIR, stem + ".wav"), wav, m.sr)
     with open(os.path.join(OUT_DIR, stem + ".srt"), "w") as f: