fix(embedding): run warmup in a worker thread to avoid blocking the event loop

warmup() can download an hf: tokenizer, which is blocking network I/O. Both call sites live in async functions (the API lifespan and run_deriver), so a slow download would stall the event loop during startup, delaying telemetry tasks and signal handling. Wrap both calls in asyncio.to_thread; warmup itself synchronizes via the singleton lock, so running it in a worker thread is safe.

src/deriver/__main__.py

@@ -73,9 +73,11 @@ async def run_deriver():
         await validate_embedding_schema(engine)
         # Eagerly build the embedding client so an hf: tokenizer download
         # happens here, not on the first reconciler call under the lock.
+        # Run it in a worker thread so the blocking download doesn't stall
+        # the event loop.
         from src.embedding_client import embedding_client
 
-        embedding_client.warmup()
+        await asyncio.to_thread(embedding_client.warmup)
         await main()
     finally:
         # Shutdown telemetry (flush CloudEvents buffer)

src/main.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 import re
 import time
@@ -118,8 +119,10 @@ async def lifespan(_: FastAPI):
     await validate_embedding_schema(engine)
 
     # Eagerly build the embedding client so an hf: tokenizer download happens
-    # here, not on the first request under the singleton lock.
-    embedding_client.warmup()
+    # here, not on the first request under the singleton lock. Run it in a
+    # worker thread: the download is blocking network I/O and would otherwise
+    # stall the event loop (signal handling, telemetry) during startup.
+    await asyncio.to_thread(embedding_client.warmup)
 
     try:
         await init_cache()