fix(loader): force re-run when keep_loaded=False so freed model gets reloaded

ComfyUI caches each node's output keyed by its inputs. With keep_loaded=False,
the previous run's Generate node freed the model and processor from the
cache (model=None, processor=None on the still-cached MOSS_TTS_MODEL handle).
Without IS_CHANGED, ComfyUI hands that empty handle to the next run's
Generate, which crashes with 'NoneType has no attribute build_user_message'.

Add IS_CHANGED that returns a fresh value (time.time()) when keep_loaded
is False, forcing ComfyUI to re-execute load() each run. get_or_load then
sees an empty cache and reloads the model from the local HuggingFace cache.

For keep_loaded=True the return value is stable, so ComfyUI keeps reusing
the same loaded model — preserves the original keep-resident behaviour.

nodes/loader.py

@@ -1,5 +1,7 @@
 """ComfyUI node: MOSSTTSLoader."""
 
+import time
+
 from ..lib.model_state import get_or_load
 
 
@@ -27,6 +29,16 @@ class MOSSTTSLoader:
     FUNCTION = "load"
     CATEGORY = "MOSS-TTS/loaders"
 
+    @classmethod
+    def IS_CHANGED(cls, model_id, device, dtype, attn_implementation, keep_loaded):
+        # When keep_loaded is False, the previous run's Generate node freed the
+        # model from the cache. ComfyUI's per-node output cache would otherwise
+        # hand the now-empty handle to Generate again. Returning a fresh value
+        # each call forces ComfyUI to re-execute load() and reload the model.
+        if not keep_loaded:
+            return time.time()
+        return f"{model_id}|{device}|{dtype}|{attn_implementation}|keep"
+
     def load(self, model_id, device, dtype, attn_implementation, keep_loaded):
         entry = get_or_load(
             model_id=model_id,