fix(core): read provider errors from plain AI SDK stream errors - #51194

Issue for this PR

N/A. Found while testing default output limits for #51021.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Provider errors that the AI SDK streams as plain objects now keep their message and classification. Before, they showed up as "[object Object]" and were retried as unknown errors.

The Vercel AI Gateway streams upstream rejections as plain objects, not Error instances. Two shapes came up live:

{ "name": "AI_TypeValidationError", "cause": { ... },
  "value": { "error": { "code": "invalid_parameter_error", "type": "invalid_request_error",
    "message": "<400> InternalError.Algo.InvalidParameter: Range of max_tokens should be [1, 32768]" } } }

{ "code": "invalid_request", "message": "Unable to submit request because This model does not support system instructions.." }

unknownErrorMessage called String(error) on these, and the fallback always built an UnknownProviderError, which the runner retries.

• Message: it is read from the object's message or error.message, including under value. When there is none, it falls back to "Provider request failed".
• Classification: the fallback goes through RequestExecutor.httpFailure, the same classifier as HTTP failures, using the object (or its value) as the error data. The first shape carries invalid_request_error, which already classifies as an invalid request.
• invalid_request: added to the invalid-request codes. It is the gateway's code for an upstream rejection.

How did you verify your code works?

• Unit test: both gateway shapes keep their message and classify as invalid requests. The packages/core AI SDK suites and the full packages/ai suite pass.

• Live through the Vercel AI Gateway, with feat(core): fit output limits to the context window #51021 applied:

  Model	Before	After
  alibaba/qwen3-max-thinking max	"[object Object]", 11 attempts	"Range of max_tokens should be [1, 32768]", 1 attempt
  google/gemini-omni-flash-preview	"[object Object]", 11 attempts	"This model does not support system instructions", 1 attempt

Screenshots / recordings

N/A

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR