Keep multimodal content inline in v1 rollouts

[eligotts]

Summary

• keep base64 image content in canonical v1 message nodes instead of carrying processed multimodal sidecars
• use renderer logical prompt tokens only for bridge extension, then store the effective expanded prompt tail returned by vLLM
• remove multimodal kwargs and metadata bookkeeping from graph, trace, and training client state
• preserve strict prefix validation and response masking across multimodal tool turns

Stack

• Renderers: Add raw multimodal content parts support renderers#133
• Prime RL: Train from inline multimodal rollouts prime-rl#3320

Validation

• uv run pytest -q tests/v1/test_graph.py (7 passed)
• uv run ruff check on all touched files

Note

Keep multimodal content inline in v1 rollouts and track logical renderer tokens

Replaces separate multi_modal_data passing with inline raw multimodal content in the prompt. Removes multi_modal_data from TurnTokens, MessageNode, Branch, and serialization logic.

• Adds renderer_prompt_ids to TurnTokens and renderer_token_ids to MessageNode to carry a separate logical renderer token stream distinct from canonical prompt tokens; both are excluded from serialization.
• TrainClient.get_response now detects raw multimodal prompts via _has_multimodal_content and requires renderer.supports_raw_multimodal; passes raw_multimodal to generate() and sets process_multimodal=False in render and bridge paths.
• PendingTurn.previous_token_ids and _commit_turn use the renderer token stream when available, and _commit_turn enforces exact prefix extension of renderer_prompt_ids against stored renderer_token_ids, raising ValueError on mismatch.
• Risk: renderers that do not implement supports_raw_multimodal now raise NotImplementedError for multimodal prompts in get_response; _commit_turn raises ValueError when incoming renderer_prompt_ids diverge from the stored rollout prefix, affecting callers that previously relied on loose prefix matching.

^{Macroscope summarized 6f5dccb.}