fix(rollout): guard round(None) in zero-std metric aggregation

[DavidBellamy]

Problem

_compute_zero_std_metrics in miles/ray/rollout.py crashes RolloutManager.generate() with a TypeError when any zero-std group's leading sample has a None reward:

ray.exceptions.RayTaskError(TypeError): ray::RolloutManager.generate()
  File "/root/miles/miles/ray/rollout.py", line 1266, in _compute_zero_std_metrics
    interesting_rewards = [str(round(g[0].get_reward_value(args), 1)) for g in interesting_sample_groups]
TypeError: type NoneType doesn't define __round__ method

This fires inside _log_rollout_data, after rollout collection and dynamic_sampling_filter have already accepted the batch. The trainer actor never receives the batch, so optimizer.step() never runs and async RL silently stalls.

Samples with Status.ABORTED carry reward=None (Daytona sandbox timeout, tool-invocation loops, agent LimitsExceeded, etc.), so any grouping of aborted trials triggers this path on GRPO-style runs.

Fix

Extract a small _reward_label helper that buckets None rewards under a dedicated "none" label instead of passing None to round(). Keeps the metric informative (zero_std/count_none shows the aborted-group count) and preserves existing behavior for numeric rewards.

Test

Smoke-checked the helper locally against numeric, float-rounding, and None cases:

assert _reward_label(MockSample(0.7)) == '0.7'
assert _reward_label(MockSample(0.0)) == '0.0'
assert _reward_label(MockSample(None)) == 'none'
assert _reward_label(MockSample(0.456)) == '0.5'

Context

Observed on LLM360/RL360#76 FAST_ITER smoke runs (GLM-4.7-Flash, agentic terminal-bench tasks) on M2. Multiple iter jobs (1559799, 1559800) reached Finish rollout successfully but crashed at this exact call site during metric aggregation. Post-fix, the trainer should receive the collected batch and fire optimizer.step() as expected.

[DavidBellamy]

Re-opening against radixark/miles:main so the LLM360/miles deploy branch auto-builder picks it up (the workflow at .github/workflows/maintain-deploy.yml queries PRs on the radixark upstream with LLM360 head branches). Same branch, same diff.