[salma2vec]

Summary

Closes #1144

gather_trajectory_groups and gather_trajectories used asyncio.gather, which re-raises the first exception but leaves the other awaitables running. When a rollout failed past max_exceptions, the caller got the error while the sibling rollouts kept hitting inference, environments and after_each judges with no owner. #1140 / #1141 cover the same gap in trajectory_group_batches.

Changes

• src/art/gather.py: a small _gather_or_cancel helper wraps the rollouts as tasks and awaits asyncio.gather. On any exception it cancels the remaining tasks, awaits them, then re-raises the original exception. The type is unchanged: TaskGroup would wrap it in an ExceptionGroup. Both gather functions use it.
• Tasks are still created inside set_gather_context, as asyncio.gather did, so they see the same context. Results, ordering, the tolerated-exception path and cancellation of the awaiting task are unchanged.
• A cancelled sibling goes through wrap_group_awaitable's existing BaseException handling, so it adds to the failed gather's exceptions tally in the progress bar. That context is discarded when the gather raises.
• tests/unit/test_gather_cleanup.py: an over-budget failure cancels the sibling in each function, and a control checks that a tolerated failure still returns every other group.

Test plan

• The two cancellation tests fail on main (5f0c1fb) and pass here. The tolerated-failure control passes on both.
• The gather_trajectory_groups / gather_trajectories leave sibling rollouts running after a failure #1144 repro prints ['finished after gather raised'] on main and [] here.
• With test_batches_completion_bursts, test_ruler_metrics, test_auto_trajectory, test_metrics_builder, test_trajectory_copy, test_yield_trajectory: 83 passed.
• ruff format / ruff check (0.15.11) clean.
• Full prek / ty not run (no CUDA backend deps on macOS).

Found and fixed with Claude Code. The runs above were executed locally.

🤖 Generated with Claude Code