Instrument Jev calls from @langchain/typesafe as evaluate spans #24859

@langchain/typesafe (TypeSafeClassifier, a LangChain Runnable) calls Jev with fetch directly. It does not use @typesafe-ai/sdk, so our TypeSafe integration does not see these calls.

Our LangChain callback handler sees the run through handleChainStart. It records the run as an invoke_agent TypeSafeClassifier span (gen_ai.invoke_agent). This span has no request model, response model, token usage, questions or answers.

Expected

An evaluate <model> span (gen_ai.evaluate) with the same attributes as the TypeSafe integration (packages/server-utils/src/ai/typesafe/): gen_ai.input.messages (state + questions), gen_ai.output.messages (answers), request/response model and token usage.

Notes

• Package: @langchain/typesafe 0.0.1, peer @langchain/core ^1.0.0. Default model jev-latest, endpoint POST {baseUrl}/v1/systemone.
• Possible approach: detect TypeSafeClassifier in handleChainStart (lc_namespace is langchain/classifiers/typesafe) and reuse the TypeSafe attribute helpers (getEvaluationInputMessages, addResponseAttributes). The chain outputs are the parsed response.
• Related: Vercel AI experimental_evaluate (feat(server-utils): Instrument Vercel AI experimental_evaluate #24694), TypeSafe integration (feat(server-utils): Add TypeSafe integration #24703), Workers AI Jev (feat(cloudflare): Trace TypeSafe Jev calls in Workers AI as evaluate spans #24833).