fix(core): retry transient response stream errors

[opencode-agent]

Summary

• decode the nested error envelope emitted by production OpenAI Responses streams
• classify known transient stream failures as retryable and stop parsing after terminal error events
• retry provider turns at most twice, and only before any assistant output is durably published
• preserve the terminal provider failure after retry exhaustion

This adapts the intent of #30323 to the current native LLM + V2 Session runner architecture.

Fixes #16214

Why

OpenAI Responses can emit transient SSE errors as:

{"type":"error","error":{"type":"server_error","code":"server_error","message":"..."}}

The parser only read top-level fields and response.error, so this became the generic OpenAI Responses stream error. The current Session runner then treated that transient stream failure as terminal, stopping agent work after completed tool calls.

Safety

Retries are bounded to two retries with exponential delay. A retry is admitted only before the Session runner has started durable assistant output, so partial text, reasoning, and tool activity are never replayed.

Verification

• packages/llm: 283 passed, 30 skipped
• packages/core: test/session-runner.test.ts — 82 passed
• bun typecheck in packages/llm
• bun typecheck in packages/core
• repository push hook: 29 package typechecks passed
• formatting, lint (0 errors), and git diff --check