[dzorlu]

Problem. An agentic episode whose environment never opened comes back ABORTED with reward None. _compute_zero_std_metrics rounds each group's reward to build its count key, so one such group raises TypeError and the whole eval pass fails.

groups: A = [None, None] (both aborted), B = [1.0, 1.0]
now:   round(None) -> TypeError, no eval metrics written
fixed: A is skipped: zero_std/count_1.0 = 1, zero_std/all_one_percentage = 0.5, no zero_std/count_None

Fix. _is_zero_std returns False for a group that has a None reward, so the group is left out of the count.

     def _is_zero_std(samples: list[Sample]):
         rewards = [sample.get_reward_value(args) for sample in samples]
+        if any(reward is None for reward in rewards):
+            return False
         return len(rewards) == 0 or all(rewards[0] == r for r in rewards)

Tests. tests/fast/ray/rollout/test_metrics.py::test_zero_std_metrics_skip_a_group_whose_episode_aborted.

Without the fix, the new test fails with TypeError: type NoneType doesn't define __round__ method at miles/ray/rollout/metrics.py:324. On CPU, tests/fast/ray/rollout has 0 new failures against main (the same 5 tests fail on both: 2 in test_engine_addressing.py, 3 in test_inference_runtime_topology.py).

Not yet proven in a run: no run so far had an aborted group reach eval scoring with this change.

🤖 Generated with Claude Code

Verified on GPU (2026-10-31)

Qwen3.8-27B VL on Megatron, 4 nodes of 8x B300, computer-use episodes through the TITO session server (screenshots, context folding on), 3 train steps with an eval pass between steps: all 3 eval passes completed (32 episodes each). Group cancel fired twice during training ("cancelled 4 running sibling episode(s)" and 7), but no eval group held a cancelled episode, so the case this PR fixes was not exercised. Not yet verified on GPU; the CPU tests cover it.