Lock YuE to a 16GB Comfy-safe recipe instead of chasing single knobs.

Keep mmgp profile 2 (no quanto), node-legal token floor, two Stage-A segments, and document why profiles 3-5 blow up under InferenceMode.

Co-authored-by: Cursor <cursoragent@cursor.com>

server/assets/workflow_yue.json

@@ -7,7 +7,7 @@
       "use_mmgp": true,
       "stage1_cache_size": 8192,
       "exllamav2_cache_mode": "FP16",
-      "mmgp_profile": 4
+      "mmgp_profile": 2
     },
     "class_type": "YUE_Stage_A_Loader",
     "_meta": {
@@ -22,7 +22,7 @@
       "run_n_segment": 2,
       "repetition_penalty": 1.1,
       "prompt_start_time": 0,
-      "prompt_end_time": 30,
+      "prompt_end_time": 45,
       "max_new_tokens": 2944,
       "use_dual_tracks_prompt": false,
       "use_audio_prompt": false,