fix: use mean().mean() for console metrics to match wandb

[mikasenghaas]

Summary

• Console reward and val reward were using results_df.reward.mean() (flat mean across all rollouts), while wandb logged by_example.reward.mean().mean() (mean per example, then across examples)
• These diverge when rollouts_per_example varies across problems (e.g. due to filtering or errors)
• Now both console and wandb use the same mean().mean() aggregation
• Also reuses the existing by_example groupby instead of re-computing it

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk: only changes how step-level console metrics are aggregated and formatted, without affecting training, logging payloads, or model behavior.

Overview
Console step logging in orchestrator.py now computes train and validation reward as mean per example, then mean across examples (groupby("example_id").mean().mean()), matching the aggregation used for W&B/monitor logging when rollout counts vary.

It also reuses the existing by_example groupby for train reward/sequence length and builds the validation string separately to simplify the step message.

^{Written by Cursor Bugbot for commit 387afac. This will update automatically on new commits. Configure here.}