fix: report failed rollout requests

- Return structured finish and error reasons for synchronous generation
- Reject failed online rollout batches instead of training on empty responses
- Verify allocation and extension failures release metrics and KV state

astrai/inference/task.py

@@ -2,8 +2,9 @@ import threading
 import time
 import uuid
 from collections import deque
+from dataclasses import dataclass
 from enum import Enum
-from typing import TYPE_CHECKING, Any, Callable, Deque, Dict, List, Optional
+from typing import TYPE_CHECKING, Any, Callable, Deque, Dict, List, Literal, Optional
 
 from tokenizers.decoders import DecodeStream
 
@@ -16,6 +17,16 @@ if TYPE_CHECKING:
 STOP = object()
 
 
+@dataclass(frozen=True)
+class GenerationResult:
+    """Structured terminal result for one synchronous generation request."""
+
+    token_ids: List[int]
+    logprobs: List[float]
+    finish_reason: Literal["stop", "length", "cancelled", "rejected"]
+    error_reason: Optional[str] = None
+
+
 class StreamDecoder:
     """Incremental decoder backed by the tokenizers library's DecodeStream.