feat: replace fake streaming with real SSE StreamGenerate

Use handler-base StreamGenerate() to publish real token-by-token
ChatStreamChunk messages to NATS as they arrive from Ray Serve,
instead of calling Generate() and splitting into 4-word chunks.

Add 8 streaming tests: happy path, system prompt, RAG context,
nil callback, timeout, HTTP error, context canceled, fallback.

main.go

@@ -157,8 +157,27 @@ func main() {
 			}
 		}
 
-		// 5. Generate LLM response
-		responseText, err := llm.Generate(ctx, query, contextText, systemPrompt)
+		// 5. Generate LLM response (streaming when requested)
+		var responseText string
+		if req.EnableStreaming {
+			streamSubject := fmt.Sprintf("ai.chat.response.stream.%s", requestID)
+			responseText, err = llm.StreamGenerate(ctx, query, contextText, systemPrompt, func(token string) {
+				_ = h.NATS.Publish(streamSubject, &messages.ChatStreamChunk{
+					RequestID: requestID,
+					Type:      "chunk",
+					Content:   token,
+					Timestamp: messages.Timestamp(),
+				})
+			})
+			_ = h.NATS.Publish(streamSubject, &messages.ChatStreamChunk{
+				RequestID: requestID,
+				Type:      "done",
+				Done:      true,
+				Timestamp: messages.Timestamp(),
+			})
+		} else {
+			responseText, err = llm.Generate(ctx, query, contextText, systemPrompt)
+		}
 		if err != nil {
 			slog.Error("LLM generation failed", "error", err)
 			return &messages.ChatResponse{
@@ -168,33 +187,7 @@ func main() {
 			}, nil
 		}
 
-		// 6. Stream chunks if requested
-		if req.EnableStreaming {
-			streamSubject := fmt.Sprintf("ai.chat.response.stream.%s", requestID)
-			words := strings.Fields(responseText)
-			chunkSize := 4
-			for i := 0; i < len(words); i += chunkSize {
-				end := i + chunkSize
-				if end > len(words) {
-					end = len(words)
-				}
-				chunk := strings.Join(words[i:end], " ")
-				_ = h.NATS.Publish(streamSubject, &messages.ChatStreamChunk{
-					RequestID: requestID,
-					Type:      "chunk",
-					Content:   chunk,
-					Timestamp: messages.Timestamp(),
-				})
-			}
-			_ = h.NATS.Publish(streamSubject, &messages.ChatStreamChunk{
-				RequestID: requestID,
-				Type:      "done",
-				Done:      true,
-				Timestamp: messages.Timestamp(),
-			})
-		}
-
-		// 7. Optional TTS — audio as raw bytes (no base64)
+		// 6. Optional TTS — audio as raw bytes (no base64)
 		var audio []byte
 		if reqEnableTTS && tts != nil {
 			audioBytes, err := tts.Synthesize(ctx, responseText, ttsLanguage, "")