fix: reduce CPU usage during large XML streaming (#801)

* fix: reduce CPU usage during large XML streaming

Skip Prism syntax highlighting while tool call is streaming — use plain
<pre> during streaming and only run Prism once at completion. Also throttle
scrollIntoView to once per 150ms to avoid layout thrashing.

Profiled with Playwright + CDP: for 200-cell diagrams the longest
browser task dropped from 6.7s to ~450ms and total long-task time
fell from ~10.9s to ~1.1s.

* fix: add trailing edge to scroll throttle

Ensures the chat scrolls to the bottom after streaming ends, even if
the last messages update arrives during the 150ms throttle window.

components/chat-message-display.tsx

@@ -417,6 +417,7 @@ export function ChatMessageDisplay({
 
     // Track previous message count to detect bulk loads vs streaming
     const prevMessageCountRef = useRef(0)
+    const scrollThrottleRef = useRef<ReturnType<typeof setTimeout> | null>(null)
 
     useEffect(() => {
         if (messagesEndRef.current && messages.length > 0) {
@@ -430,8 +431,17 @@ export function ChatMessageDisplay({
                 return
             }
 
-            // Single message added - smooth scroll
-            messagesEndRef.current.scrollIntoView({ behavior: "smooth" })
+            // Throttle scroll during streaming to avoid layout thrashing
+            // Leading + trailing: scroll immediately, then once more after cooldown
+            if (!scrollThrottleRef.current) {
+                messagesEndRef.current.scrollIntoView({ behavior: "smooth" })
+                scrollThrottleRef.current = setTimeout(() => {
+                    scrollThrottleRef.current = null
+                    messagesEndRef.current?.scrollIntoView({
+                        behavior: "smooth",
+                    })
+                }, 150)
+            }
         }
     }, [messages])