abort(): register live socket at urllib3 request() so Force Stop wakes recv() even while waiting for response headers (was only after headers via active_response)

agentmain.py

@@ -116,12 +116,10 @@ class GenericAgent:
         if self.handler is not None: self.handler.code_stop_signal.append(1)
         for sess in getattr(self.llmclient.backend, '_sessions', [self.llmclient.backend]):
             sess.should_stop = lambda: self.stop_sig  # live read; cleared by run()'s finally
-            try:  # wake a recv() blocked in another thread. Verified on Windows: shutdown()/close() do NOT
-                  # wake it (makefile refcount defers real closesocket); _real_close() does -> ChunkedEncodingError
+            try:  # wake a recv() blocked in another thread (waiting headers OR reading stream). Verified on Windows:
+                  # shutdown()/close() do NOT wake it (makefile refcount defers closesocket); _real_close() does
                 import socket as _socket
-                raw = sess.active_response.raw
-                fp = getattr(getattr(raw, '_fp', None), 'fp', None)  # http.client response -> buffered socket file
-                sock = fp.raw._sock if fp else raw.connection.sock   # SocketIO._sock (SSL-wrapped OK); fallback urllib3 conn
+                sock = sys.modules['llmcore']._INFLIGHT[sess._tid]  # socket registered at urllib3 request() time, before headers
                 try: sock.shutdown(_socket.SHUT_RDWR)  # for non-Windows semantics
                 except OSError: pass
                 try: sock._real_close()  # CPython internal; bypasses refcount -> actual closesocket