v0.3.0

Released April 2, 2026.

• Added a thin native DAPO adapter with reward/config helpers for verl-style training recipes
• Hardened imported benchmark loading so nested HumanEval/MBPP-style benchmark registries resolve predictably
• Fixed benchmark alias and task-id routing for imported suites such as humaneval and HumanEval/0
• Made mixed benchmark routing stricter and rejected mismatched per-sample batch metadata instead of silently broadcasting it
• Restored Python 3.10 compatibility in the release/test matrix