Instrument TypeSafe Jev (@typesafe-ai/sdk) #24693

TypeSafe's official SDK (@typesafe-ai/sdk, latest 0.6.0) calls the Jev evaluation model with TypeSafeClient.systemOne({ state, questions, model? }). Users who call Jev directly (not through Vercel AI experimental_evaluate, see #24692) get no span today, only a plain http.client span from fetch.

Scope

• Add a new AI integration in packages/server-utils/: logic in src/ai/typesafe/, integration in src/integrations/typesafe.ts, registered in getTracingIntegrations(). Re-export it from the runtime packages.
• Instrument TypeSafeClient.prototype.systemOne. Make one span for each call, not one span for each retry.
• models.list() is out of scope.

Target span

Use the same shape as the Vercel AI evaluate span in #24692, so that both paths look the same in the product.

• Name: evaluate <model>, op gen_ai.evaluate, origin auto.ai.typesafe
• gen_ai.operation.name: evaluate, gen_ai.provider.name: typesafe
• gen_ai.request.model (the model option, or the client default, jev-latest), gen_ai.response.model (result.model, for example jev-1.13.0)
• gen_ai.usage.input_tokens, output_tokens, total_tokens (from result.usage.input_tokens / output_tokens)
• gen_ai.input.messages: [{ type: 'evaluation', state, questions }] (only when recordInputs is on)
• gen_ai.output.messages: [{ type: 'evaluation', answers }] (only when recordOutputs is on)
• Set the span status to error when the call throws (TypeSafeError and API errors).

Notes

• The SDK sends requests with fetch to POST /v1/systemone. It has no tracing channel, so we must patch the client.
• @sentry/conventions has no gen_ai.evaluate op yet (same gap as Instrument Vercel AI experimental_evaluate #24692).
• Add a node-integration test with a mocked TypeSafe API server.