fix: re-schedule empty trajectories to guarantee full groups

[mikasenghaas]

Summary

• Empty trajectories were filtered after group completion and buffer sampling, which could yield groups with fewer than rollouts_per_example rollouts
• This broke the advantage computation which reshapes rewards with .view(-1, rollouts_per_example) — either crashing on misaligned tensors or silently computing wrong per-problem baselines
• Now empty trajectories are detected per-rollout as they arrive. When one is found, rollouts_to_schedule is incremented so the group naturally re-fills, and only complete groups are yielded

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches core rollout scheduling/group completion logic; a bug could cause groups to stall or skew batch composition, but the change is small and localized.

Overview
Prevents incomplete rollout groups by handling empty trajectories before group completion.

In Scheduler.generate_batch, each finished rollout is now checked for an empty trajectory; empty results increment batch/empty_rollouts, log a warning, and increase group.rollouts_to_schedule so the group naturally re-fills, while the previous post-sampling filtering of empty rollouts is removed.

^{Written by Cursor Bugbot for commit b43629f. This will update automatically on new commits. Configure here.}