fix: count unscored reward components as zero in rollout metrics

[mikasenghaas]

Summary

• Companion to feat: seed unscored rewards/metrics as None placeholders verifiers#2235: scoring now seeds every expected reward/metric name with None before invoking it, so a scoring failure leaves seeded-unscored entries ({"solved": null}) instead of an empty dict.
• CustomMetrics counts unscored (None) reward and metric entries as 0.0. Previously scoring-failed traces silently dropped out of all/rewards/<name>/mean and all/metrics/<name>/mean, making the all subset collapse onto effective — while all/reward/mean counted the same traces as 0, so the views couldn't be reconciled. Now all is the raw view including failures as zeros, and effective (which excludes errored rollouts) is the clean view.
• Cross-env agg pools need no special handling: another env's unscored traces carry their own env's keys, so they can't dilute a different env's means.
• Bumps deps/verifiers to the seeding commit.

Uncovered on scaleswe_v1 SWE-bench Verified evals. Linear: RES-1210

Breaking

• {train,eval}/*/all/{rewards,metrics}/<name>/* wandb series change meaning: traces whose scoring failed after seeding now count as 0.0 for each expected key (previously they were silently excluded, so all matched effective). effective subsets are unchanged.

Verification

uv run pytest tests/unit passes against the bumped submodule; test_nested_metrics_and_rewards covers the new semantics (unscored entries → 0.0 in all, dropped from effective; cross-env isolation).

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes the meaning of existing all/{metrics,rewards}/<name> wandb series (breaking for dashboards); logic is localized to aggregation but affects how eval/train metrics are interpreted.

Overview
CustomMetrics now treats seeded-but-unscored entries (None in per-rollout metrics / rewards dicts) as 0.0 when building per-key means, instead of skipping those rollouts for that key.

That makes the all wandb slice include scoring-failed traces as zeros for each expected component, aligning with scalar all/reward/mean behavior. The effective subset is unchanged (errored rollouts are still excluded there).

Unit tests in test_nested_metrics_and_rewards were extended for errored rollouts with None seeds, effective means, and cross-env aggregation so unrelated env keys do not dilute another env’s component means.

^{Reviewed by Cursor Bugbot for commit bc41c58. Bugbot is set up for automated code reviews on this repo. Configure here.}