feat: preserve token IDs on messages

[Hecate0821]

Summary

Adds token-native trace support to EP messages without turning EP into an RL framework.

Current Design

eval_protocol.models.Message now has optional token_ids: list[int] | None. This lets rollout processors preserve engine token IDs alongside message content and provider logprob metadata.

The schema remains backward compatible:

• Provider-specific logprobs dictionaries are still accepted unchanged.
• Text-only messages still work.
• A strict alignment check only applies to the clean token-native shape where both token_ids and flat list[float] logprobs are set. In that case, lengths must match.

SingleTurnRolloutProcessor now extracts token IDs from serialized provider logprobs when they are available:

• logprobs.content[].token_id
• logprobs.token_ids[]

Those IDs are stored on the assistant Message.token_ids. If the provider does not expose token IDs, the field remains unset.

Why

The cookbook async RL path is token-native only. Multi-turn RL cannot safely re-tokenize assistant text after rollout because BPE merges can cross turn boundaries and inference logprobs can become misaligned. EP needs a simple message-level carrier for token IDs so downstream training can consume traces without reconstructing tokens from text.

Tests / Checks

• python3.11 -m pytest tests/test_rollout_logprobs.py tests/test_eval_protocol_import.py::TestRewardProtocolFunctionality::test_message_preserves_token_ids tests/test_eval_protocol_import.py::TestRewardProtocolFunctionality::test_message_rejects_misaligned_float_logprobs
• python3.11 -m ruff check eval_protocol/models.py eval_protocol/pytest/default_single_turn_rollout_process.py tests/test_rollout_logprobs.py tests/test_eval_protocol_import.py
• git diff --check

Note: full tests/test_eval_protocol_import.py currently has an unrelated failure in TestRewardProtocolImports.test_star_import_works because eval_protocol.adapters does not expose LangfuseAdapter.

Related

• Design issue: fw-ai/fireworks#23756
• Cookbook PR: feat(rl): gate-native async RL loop + flat Rollout contract fw-ai/cookbook#382