ai: classify plain AI SDK stream errors #43791

Summary

Plain errors emitted by an AI SDK model stream bypass shared provider-failure classification. Transient messages such as provider capacity errors are therefore classified as UnknownProvider and are not retried, while equivalent APICallError and native AI-package failures use classifyProviderFailure correctly.

Environment

• opencode version: v0.0.0-dev-17773
• OS: Darwin 25.6.0 arm64
• Terminal: Ghostty (TERM=xterm-256color, COLORTERM=truecolor)
• Shell: /bin/zsh
• Install/channel: dev, V2
• Active plugins: none configured

Reproduction

1. Configure an AI SDK language model whose stream emits an error part containing a plain Error, rather than an APICallError.
2. Use a transient provider message such as The model is currently at capacity due to high demand. Please try again in a few minutes.
3. Run a prompt through the V2 Session runner.

Expected Behavior

The plain stream error message should pass through the shared provider-failure classifier. Known transient text should become ProviderInternal and participate in the existing retry policy.

Actual Behavior

packages/core/src/aisdk.ts routes APICallError through RequestExecutor.classifyHttpFailure, which reaches classifyProviderFailure. A plain Error instead falls through llmError() directly to UnknownProviderReason, so transient text matching never runs and the failure is not retried.

Additional Context

This affects AI SDK stream errors from language.doStream(), readable-stream failures, and LanguageModelV3StreamPart error events whenever the underlying value is a plain error. Native AI-package HTTP failures and modeled stream error events already converge on packages/ai/src/provider-error.ts.

The fix should preserve unknown plain errors as UnknownProvider while allowing known structured or textual classifications to use the shared classifier. Add coverage for a plain AI SDK stream error carrying transient capacity text, alongside the existing APICallError tests.