Add exclusive LLM and video modes with isolated LTX worker and job API

inference.py

@@ -19,7 +19,7 @@ class InferenceError(ValueError):
 class Scheduler:
     """FIFO admission; same-profile requests share configured slots, switches drain."""
     def __init__(self, worker):
-        self.worker=worker;self.cv=threading.Condition();self.queue=[];self.key=None;self.active=0;self.transition=False
+        self.worker=worker;self.cv=threading.Condition();self.queue=[];self.key=None;self.active=0;self.transition=False;self.gpu_mode="llm"
     def status(self):
         with self.cv:return dict(active_requests=self.active,waiting_requests=len(self.queue),switching=self.transition)
     @contextlib.contextmanager
@@ -30,6 +30,7 @@ class Scheduler:
             self.queue.append(ticket)
             try:
                 while True:
+                    if self.gpu_mode!="llm":raise InferenceError("Video-Modus aktiv oder Moduswechsel läuft; GPU-Aufträge sind gesperrt.")
                     if not allowed():raise InferenceError('Endpunkt wird gestoppt oder Profil ist nicht mehr aktiviert.')
                     if self.queue[0] is ticket and not self.transition and (not self.active or (self.key==key and self.active<slots)):
                         self.queue.pop(0);self.active+=1;claimed=True