fix(legacy): preserve chat prompt roles

[yaodong-shen]

Summary

• preserve structured chat messages during token-budget truncation for the legacy OpenAI, MiniMax, and VLLM API generators
• send the complete role-preserving message list to the GPT-5 Responses API instead of flattening system/user turns and dropping assistant history
• retain the existing MiniMax helper import surface while sharing the truncation implementation

Root cause

The OpenAI and VLLM API paths converted List[Dict] chat prompts to a single string before dispatch. That turned role markers into user content; GPT-5 additionally discarded prior assistant turns when forming a Responses request.

Validation

• PYTHONPATH=/private/tmp/autorag-test-stubs uv run --no-sync pytest tests/autorag/nodes/generator/test_openai.py tests/autorag/nodes/generator/test_minimax.py tests/autorag/nodes/generator/test_vllm_api.py -q -k 'truncate_chat_prompt_preserves_message_roles or gpt_5_responses_preserve_full_chat_history or truncate_messages_preserves_roles or vllm_api_truncate_preserves_chat_messages'
  • 4 passed
• uvx ruff check on all changed generator and test files
• uvx ruff format --check on all changed generator and test files
• git diff --check

The full legacy suite was not runnable locally because its macOS ARM development resolution includes an NVIDIA-only package and its OpenAI test module imports optional torch.cuda during collection. The focused regression tests use no model call, credentials, or GPU.

Closes #1234