fix(studio): forward AI thinking chunks to frontend instead of dropping them

The ThinkingBlock component existed but was dead code — the backend
silently discarded all <think chunks. Now emits thinking SSE events
so the UI can display AI reflections in real-time.

\xe2\x98\x85 Generated with Crush

Assisted-by: GLM-5-Turbo via Crush <crush@charm.land>

internal/api/handlers_chat.go

@@ -73,8 +73,20 @@ RÈGLES ABSOLUES:
 		flusher, canFlush := w.(http.Flusher)
 
 		result, err := orb.SendStream(body.Message, func(chunk string) {
-			// Skip thinking tags - user doesn't see them
 			if strings.HasPrefix(chunk, "<think") {
+				data, _ := json.Marshal(map[string]string{"thinking": strings.TrimPrefix(chunk, "<think")})
+				w.Write([]byte("data: " + string(data) + "\n\n"))
+				if canFlush {
+					flusher.Flush()
+				}
+				return
+			}
+			if chunk == "</think>" {
+				data, _ := json.Marshal(map[string]string{"thinking_end": "true"})
+				w.Write([]byte("data: " + string(data) + "\n\n"))
+				if canFlush {
+					flusher.Flush()
+				}
 				return
 			}
 			data, _ := json.Marshal(map[string]string{"content": chunk})

web/src/api/client.js

@@ -68,7 +68,9 @@ const api = {
               if (data.done) { resolve(full); return }
               if (data.content) {
                 full += data.content
-                if (onChunk) onChunk(full)
+                if (onChunk) onChunk(full, data)
+              } else if (data.thinking !== undefined || data.thinking_end) {
+                if (onChunk) onChunk(full, data)
               }
             } catch {}
           }