refactor: unify rollout onto inference engine KV-cache path

- RolloutGenerator now delegates prefill/decode to InferenceScheduler.run_batch (sync API, no background thread), sharing one KV-cache code path with the inference server and eliminating O(n^2) recompute in rollout
- Add sample(return_logprobs=) and Executor.execute_decode(return_logprobs=) to expose behaviour-policy log-probs through the engine; Task gains output_logprobs
- RolloutResult now subclasses RawRollout (adds rewards only), removing duplicated fields
- RolloutRunner.__call__ returns (result, is_fresh) instead of relying on object identity, removing the fragile refresh-detection contract
- Remove O(n^2) generate_responses helper and dead code (_tokenize_prompts, unused old_model arg)
- train_context.py wires InferenceScheduler directly instead of hand-rolling SamplingPipeline
- Tests: +11 covering return_logprobs, run_batch, and KV-cache-backed rollout semantics; 404 pass

astrai/inference/core/executor.py

@@ -55,7 +55,23 @@ class Executor:
                 paged_cache=self.kv_cache.bind_tasks(task_ids, prompt_len, self.device),
             )
 
-    def execute_decode(self, tasks: List[Task]) -> List[int]:
+    def execute_decode(
+        self, tasks: List[Task], return_logprobs: bool = False
+    ) -> List[int]:
+        """Decode next token for each task.
+
+        Args:
+            return_logprobs: When ``True``, also record (and return)
+                the log-probability of each sampled token under the
+                post-strategy sampling distribution.  The logprob is
+                appended to ``task.output_logprobs`` and the return
+                list becomes ``List[Tuple[int, float]]``.
+
+        Returns:
+            ``List[int]`` of sampled token IDs, or
+            ``List[Tuple[int, float]]`` of ``(token_id, logprob)`` when
+            ``return_logprobs`` is ``True``.
+        """
         if not tasks:
             return []
 
@@ -116,6 +132,23 @@ class Executor:
             )
             logits = outputs["logits"][:, -1, :]
 
+        if return_logprobs:
+            tokens, logprobs = sample(
+                logits,
+                temperature=temperatures,
+                top_k=top_ks,
+                top_p=top_ps,
+                frequency_penalty=freq_penalties,
+                input_ids=padded_ids,
+                input_mask=padded_mask,
+                return_logprobs=True,
+            )
+            tokens_list = tokens.tolist()
+            logprobs_list = logprobs.tolist()
+            for t, lp in zip(tasks, logprobs_list):
+                t.output_logprobs.append(float(lp))
+            return list(zip(tokens_list, logprobs_list))
+
         return sample(
             logits,
             temperature=temperatures,