feat(evaluator): aggregate agent-eval results natively and import Gym's own

[SandyChapman]

Stacked on #1013. Base is aalgo-451-run-metadata/schapman, so the diff here is only this change. Retarget to main once #1013 merges.

Closes AALGO-434.

Why

Agent-eval reported a mean and little else, and a runner's own aggregations were discarded entirely — so a Gym run's headline numbers (pass@k, per-environment metrics) were visible in Gym's output files but nowhere in the eval result.

Native aggregation

• pass@k per task via the unbiased estimator (Chen et al. 2021), for every score-like metric output. An attempt passes on full credit — deliberately not configurable, since it's a reporting-time interpretation and a tunable version would produce numbers that look comparable across runs but aren't.
• Percentiles on every aggregate, reusing the deterministic-metric helper (now public as compute_percentiles) so both paths report distributions the same way.
• median added to AggregateScoreBase, populated natively from p50, so the field means the same thing whether a score was computed here or imported from a backend that reports a median without a full distribution.
• Both standard-deviation conventions, named explicitly. std_dev/variance remain population (÷n); sample_std_dev/sample_variance are new (÷n−1, None when n<2). Gym computes its spread with pandas (ddof=1), so naming both lets the two sets of numbers coexist without either silently changing meaning. No existing value changes.

Importing Gym's own numbers

RunAggregationsProvider is a single method — run_aggregate_scores() — returning typed aggregates that merge into summary.scores. There is no parallel opaque passthrough: Gym already writes its raw rollouts_aggregate_metrics.json inside the run's work dir, so a second copy on the result would duplicate a file already in the bundle and re-introduce the untyped dict[str, Any] bag that #1013 removes.

Names are runner.gym.<metric> — namespaced by runner, not agent, since each run uses a single agent and what a reader needs to know is which backend produced the number. Falls back to runner.gym.<agent>.<metric> only if a run produced several. Gym's reward is skipped: the SDK already scores it natively as gym_reward.reward from the same rollouts.

The part most worth reviewing

Imports read agent_metrics, not key_metrics. This matters: key_metrics is a subset of agent_metrics chosen by the resources-server, and Gym's default get_key_metrics keeps only the mean/* entries — so max/min/median/std never appear there, and sourcing from it would degrade every distribution into a lone runner.gym.mean/<name> scalar. test_imports_read_agent_metrics_not_the_key_metrics_subset pins this.

Gym flattens each distribution into <stat>/<metric> keys. Its describe_dataframe emits mean/max/min/median/std/histogram together, and prepare_for_serialization strips histogram before writing — so the family in the file is exactly those five, and re-assembly requires all five. That guard matters because 36 of Gym's ~97 resources-servers override compute_metrics, emitting keys in their own shapes (arena_elo/score, easy/pass@1/accuracy), and a server is free to define a metric literally named mean. A partial match stays as standalone scalars rather than being renamed into a statistic of a distribution that never existed.

test_nothing_numeric_is_dropped_from_a_custom_environment_payload pins the invariant: importing is a renaming, never a filter. Only reward (redundant by construction) and non-numeric values may disappear.

New AggregateScalarScore

A single pre-computed value with no underlying sample. Distinct from AggregateRangeScore so a reader can tell "this is the whole story" from "this summarizes count samples", instead of seeing a range score with a suspicious count of 1.

count is now int | None, and imported aggregates carry None — Gym reports statistics without the n behind them, and 0 would assert that nothing was evaluated (besides being a division hazard). The dashboard and example render None as an em dash while still showing a genuine 0.

Also

• Dashboard gains Value / Median / Std dev columns and renders scalars by value.
• New examples/gym/inspect_results.py — companion to run_gym_eval.py, showing how to reach each kind of result, with lift-and-paste accessors.

Known scope boundary

Gym's group_level_metrics (its per-task aggregation) is not imported. summary.scores is run-level, so it has no home there yet — the SDK computes per-task groupings internally for pass@k but doesn't surface that granularity. Native per-task rollups plus importing Gym's into them is tracked separately; the raw data remains in Gym's own file inside the bundle.

Verification

• Full run: 1820 passed / 25 failed, all 25 the pre-existing ragas + live-Fabric baseline (verified: zero failures outside it).
• Ruff clean; ty at the branch base's baseline (151); make vendor run with no drift.