Group binary agent outcomes for standard metrics

[sylvesterkaczmarek]

Summary

• group agent-summary answers by their binary correctness outcome instead of assigning every rollout a unique answer group
• deduplicate the per-instance correctness mapping for the two binary groups
• preserve existing outcome counts and system-error handling

get_summary_error_aware() only has a binary result for each rollout: correct or incorrect. _compute_metrics_plus_outcome_aggregations() currently assigns answer_group_id with groupby(...).cumcount(), so every attempt is treated as a different answer even when multiple attempts have the same outcome.

That does not affect the repository's minimal accuracy-only backend, but it breaks the answer-group contract for richer standard metric backends that use group identity for consensus-style metrics. Two correct attempts should belong to the same answer group, as should two incorrect attempts.

The fix uses 1 for correct and 0 for incorrect, and keeps one correctness row per (instance, answer_group_id) so repeated outcomes do not duplicate the correctness mapping.

Regression coverage verifies that repeated correct attempts share one group while incorrect attempts use the other group.