Add ART and Verifiers trajectory interop helpers

[nboufeld]

Summary

• adds lightweight ART <-> Verifiers rollout portability helpers
• converts ART Trajectory objects into Verifiers-style rollout state with per-choice trajectory steps
• converts Verifiers rollout states back into ART Trajectory / TrajectoryGroup objects, with an optional choice_factory for preserving trainable assistant choices when callers have the original response payload
• documents the helpers in the README

Bounty fit

This targets the hard scope from the Algora ART-E / ART bounty: two-way portability between ART and Verifiers tasks.

Bounty: https://algora.io/PrimeIntellect-ai/bounties/cxptwFoEv2up9Qjm

Verification

• PYTHONPATH=src /tmp/art-interop-venv/bin/python -m pytest tests/unit/test_verifiers_interop.py -q
• /tmp/art-interop-venv/bin/ruff format src/art/verifiers.py tests/unit/test_verifiers_interop.py
• /tmp/art-interop-venv/bin/ruff check src/art/verifiers.py tests/unit/test_verifiers_interop.py
• git diff --check

Note: I used a minimal local venv for this pure-Python adapter test because the repo's current uv config attempted to resolve/build heavy optional GPU dependencies in this environment.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bb90893ea5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve completion tail when rebuilding from trajectory steps

When state["trajectory"] is non-empty, this branch returns _steps_to_messages_and_choices(...) and never merges state["completion"]. In states that include completion messages after the last recorded step (for example, environment/tool turns appended at rollout end), those tail messages are silently dropped on import, so ART↔Verifiers round-trips lose transcript content.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Propagate truncation status to state-level flag

This sets is_truncated to False unconditionally, even when imported choices mark a truncated finish (e.g., finish_reason == "length" on the final step). Consumers that read only the state-level truncation flag will misclassify truncated rollouts as complete, which can affect filtering and reward/metrics analysis.

Useful? React with 👍 / 👎.

[nboufeld]

Updated after the automated review feedback. The conversion now propagates state-level is_truncated when any imported choice/step is truncated, and verifiers_state_to_trajectory preserves state-level completion tail messages that appear after the last recorded trajectory step.\n\nRe-verified locally:\n- PYTHONPATH=src /tmp/art-interop-venv/bin/python -m pytest tests/unit/test_verifiers_interop.py -q -> 5 passed\n- /tmp/art-interop-venv/bin/ruff check src/art/verifiers.py tests/unit/test_verifiers_interop.py -> passed\n- git diff --check -> passed