fix : handle zero-token batch generation

- return empty results without running inference for non-positive limits
- keep scheduler batch outputs aligned with requested max_tokens
- add engine and scheduler regression coverage

astrai/inference/core/scheduler.py

@@ -288,6 +288,9 @@ class InferenceScheduler:
                 t_max = seq_cap - len(ids)
             else:
                 t_max = min(t_max, seq_cap - len(ids))
+            if t_max <= 0:
+                tasks.append(None)
+                continue
             task = Task(
                 task_id=f"batch_{uuid.uuid4().hex[:8]}",
                 prompt_ids=list(ids),

astrai/inference/engine.py

@@ -146,6 +146,12 @@ class InferenceEngine:
         is_batch = isinstance(prompt, list)
         prompts = prompt if is_batch else [prompt]
 
+        if max_tokens is not None and max_tokens <= 0:
+            if stream:
+                return iter(())
+            results = [""] * len(prompts)
+            return results if is_batch else results[0]
+
         if stream:
             return self._generate_streaming(
                 prompts,