fix(llm_gateway): buffer non-streaming response body until end_of_stream

Wait for the full upstream body before JSON parsing to avoid truncated
responses on chunked replies. Retry currency_exchange demo tests on flake.

Co-authored-by: Musa <musa@spherrrical.dev>

crates/llm_gateway/src/stream_context.rs

@@ -1183,6 +1183,12 @@ impl HttpContext for StreamContext {
             return Action::Continue;
         }
 
+        // Non-streaming upstream responses may arrive in multiple chunks; wait for the
+        // full buffered body before parsing (body_size is cumulative on the final chunk).
+        if !self.streaming_response && !end_of_stream {
+            return Action::Continue;
+        }
+
         // Check if this is an error response from upstream
         if let Some(status_code) = &self.upstream_status_code {
             if status_code.is_client_error() || status_code.is_server_error() {

demos/shared/test_runner/run_demo_tests.sh

@@ -19,7 +19,8 @@ run_hurl_with_retries() {
   local max_attempts=1
   local attempt=1
 
-  if [ "$demo_name" = "llm_routing/preference_based_routing" ]; then
+  if [ "$demo_name" = "llm_routing/preference_based_routing" ] \
+    || [ "$demo_name" = "advanced/currency_exchange" ]; then
     max_attempts=3
   fi