fix(error): fall back to full response body when error field is absent

[s-zx]

Summary

Fixes #1734

When an OpenAI-compatible API returns an error response that does not include an error field (e.g. using detail, message, or other custom fields at the top level), the Node.js client produced unhelpful error messages like 422 status code (no body).

This is because APIError.generate() extracted only body.error and passed undefined to the error constructors when that field was missing — the makeMessage logic then had nothing to display.

Change

In src/core/error.ts, update the error extraction in APIError.generate() to fall back to the full response body when body.error is not present:

// Before
const error = (errorResponse as Record<string, any>)?.['error'];

// After
const error = (errorResponse as Record<string, any>)?.['error'] ?? errorResponse;

This mirrors the behavior of the Python client, which falls back to including the entire response body in the error message when no error field is found. The existing makeMessage logic already handles this correctly by calling JSON.stringify() on non-string error values.

Example

With a response body like {"detail": "422: The model gpt-5-gibberish does not exist."}:

• Before: 422 status code (no body)
• After: 422 {"detail":"422: The model gpt-5-gibberish does not exist."}