Mark turn-budget exhaustion as truncated

[nightlessbaron]

Summary

• interpret Harbor exit_status=LimitsExceeded as rollout truncation
• mark only the final sample TRUNCATED for multi-sample trajectories
• retain agent metadata and all earlier completed turns

Validation

• focused helper tests: 5 passed
• added end-to-end generation coverage for merged and multi-sample variants
• Ruff and isort passed
• Black passed on the changed source, new test, and changed integration-test range
• git diff --check passed

The full integration case could not collect locally because its fixture attempted an uncached Hugging Face tokenizer lookup in the network-restricted environment; CI will run it in the repository environment.

[flukeskywalker]

It seems merge_samples is not imported, according to CI