[v5.0] reproduction: streamText ends the ai.streamText root span twice when

[ai-sdk-factory]

Bug

streamText ends the ai.streamText root span twice when the model call fails (recordSpan ignores endWhenDone on error)

Reproduction

• Target release-v5.0 contains ai@5.0.228 and the reported telemetry APIs, so the failure also applies to this release line.
• examples/ai-functions/src/reproduction/issue-18574.ts reproduced the primary bug: the failed stream emitted start,error, exported the root span before stream consumption, and produced one duplicate-end diagnostic.
• The exported ai.streamText span lacked ai.response.finishReason; on this branch, the zero-step flush also returns before the finish-attribute block, so premature ending is not conclusively the sole cause of that missing attribute.
• packages/ai/src/telemetry/record-span.ts ends rejected spans unconditionally, while packages/ai/src/generate-text/stream-text.ts passes endWhenDone: false and later ends the same root span from flush().
• Expected behavior is one root-span end after stream processing, no ended-span diagnostic, and retained applicable finish telemetry.
• examples/ai-functions/package.json and pnpm-lock.yaml provide the branch-aligned AI SDK and OpenTelemetry dependencies needed by the reproduction.

Relevant Documentation

• https://ai-sdk.dev/docs/ai-sdk-core/telemetry
• https://ai-sdk.dev/docs/reference/ai-sdk-core/stream-text
• https://opentelemetry.io/docs/specs/otel/trace/api/

Related Issues

Reproduces #18574