[ErlisLushtaku]

Fixes two vLLM problems that judge tuning runs into.

• vLLM 0.25 imports responses.NamespaceTool from the OpenAI SDK, so the vllm extra now requires openai>=2.30.0.
• One prompt longer than the model context made LLM.chat() fail the whole batch. Each chat request is now length-checked first, and overlong ones come back as failed annotations (error="context_length") while the valid requests still run. A search will hit this on long conversations, and one bad battle should not fail a whole trial. The check uses a private vLLM method (_preprocess_chat_one), which exists in 0.19 to 0.25.
• Failed requests are not written to the inference cache, and they are left out of usage accounting. That lets the cost objective in Judge tuning (4/5): Run multi-fidelity judge searches #133 tell a skipped request apart from a successful response with missing usage.

Merge order: #138 → #135 → #140 → #133 → #136.

[kargibora]

Are we sure the error is context length?

[ErlisLushtaku]

Yes, the None comes from _run_raw_batch, where we leave an empty slot for requests rejected for exceeding the context limit and fill in the results for the valid ones. Other errors are raised there, so they don't reach this loop.

[kargibora]

This doesnt break anything right?

[ErlisLushtaku]

Seems fine with the current dependencies. I tried a clean install with both extras. Also the tests pass. This resolved to vllm==0.32.0 and openai==3.35.1.