Add eval sample logging to monitor backends

[bhavishya-pohani]

Problem

Evaluation rollouts are computed and scored during training, but only aggregate metrics (avg@k, pass@k) are logged. Individual eval traces are discarded.

Training logs 8 samples per step via monitor.log_samples(), but there's no equivalent call in evaluate_env() — it only calls monitor.log(eval_metrics, ...).

Why this matters

Debugging a degrading run (reward collapse, truncation spikes, content bloat) requires inspecting what the model actually generated, not just aggregate numbers. Eval sets are small by design (30-100 examples) and run infrequently, so storage cost is minimal. The data is already being generated — it's just thrown away.

There is currently no way to access individual eval traces — not via CLI, API, or the platform UI.

Proposal

1. Eval samples: Call monitor.log_samples() for eval rollouts in evaluate_env(), the same way training does. Since eval sets are small, all samples could be logged (not a subset). A config option like log_samples = true under [eval] would work if there are storage concerns.

2. Training samples: It would also be nice to make the training rollout sample count configurable rather than hardcoded to 8 (min(8, len(train_rollouts)) in the orchestrator). For debugging, being able to download a larger subset — or all rollouts — from a training step would be very helpful.

Both would make traces available via prime rl rollouts and W&B, enabling proper post-hoc debugging of hosted runs.

[mikasenghaas]

yes, very fair. we will add this!