Fix #38: enable streaming proxy with async chunk iteration

Pass stream=True to requests.request in signed_call.py so the response
body is not buffered before being returned. Replace iter_content() in
call_backend.py with an async generator that reads 8192-byte chunks and
offloads each blocking socket read to a thread via asyncio.to_thread,
eliminating ~28s of unnecessary latency.

Author: ClaydeCode

shard_core/service/signed_call.py

@@ -15,7 +15,9 @@ async def signed_request(
     *args, identity: Identity = None, **kwargs
 ) -> requests.Response:
     auth = get_signature_auth(identity)
-    response = await asyncio.to_thread(requests.request, *args, auth=auth, **kwargs)
+    response = await asyncio.to_thread(
+        requests.request, *args, auth=auth, stream=True, **kwargs
+    )
     return response
 
 

shard_core/web/internal/call_backend.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 from fastapi import APIRouter, Request
 
@@ -11,6 +12,15 @@
 router = APIRouter()
 
 
+async def async_iter_content(response, chunk_size=8192):
+    iterator = response.iter_content(chunk_size=chunk_size)
+    while True:
+        chunk = await asyncio.to_thread(next, iterator, None)
+        if chunk is None:
+            break
+        yield chunk
+
+
 @router.api_route("/call_backend/{rest:path}", methods=ALL_HTTP_METHODS)
 async def call_backend(rest: str, request: Request):
     base_url = settings().freeshard_controller.base_url
@@ -26,5 +36,5 @@ async def call_backend(rest: str, request: Request):
     return StreamingResponse(
         status_code=response.status_code,
         headers=response.headers,
-        content=response.iter_content(),
+        content=async_iter_content(response, chunk_size=8192),
     )