fix(ai): preserve stream transport failures - #41779

Summary

• add an executor-owned typed byte-stream adapter without changing the public RequestExecutor.Interface
• decode native transport failure metadata once at the HTTP I/O boundary and preserve nested native codes and messages
• classify transport failures with stable transport and operation domain fields plus an open-ended native code
• preserve the request actually sent through middleware and keep request diagnostics redacted
• keep HTTP transports focused on framing an already typed stream
• let the existing V2 runner retry transport failures when no model output has started, while retaining the existing no-replay behavior after partial output

Verification

• bun typecheck (packages/ai)
• bun typecheck (packages/core)
• bun test test/executor.test.ts
• bun test test/provider/openai-chat.test.ts --test-name-pattern "surfaces transport errors"
• bun test test/provider/openai-responses.test.ts --test-name-pattern "streams OpenAI Responses over WebSocket|fails immediately when WebSocket is already closed"
• bun test test/aisdk.test.ts test/session-error.test.ts test/session-execution.test.ts test/session-runner.test.ts
• full monorepo typecheck from the pre-push hook

Requested by: @rekram1-node (Aiden Cline via Slack)