Add native OpenAI runtime opt-in

[kitlangton]

Summary

• add OPENCODE_LLM_RUNTIME=native as an experimental session runtime selector
• route no-tools OpenAI requests through native @opencode-ai/llm streaming while keeping AI SDK as the default
• carry opencode provider auth/base URL/headers into native requests and cover the path with a local mocked OpenAI Responses stream test

Notes

• native runtime is intentionally limited to OpenAI no-tools for this first slice
• tool-enabled requests fail clearly instead of falling back silently

Test Plan

• bun run test -- test/session/llm-native.test.ts test/session/llm.test.ts --test-name-pattern native
• bun typecheck
• bunx oxlint packages/opencode/src/session/llm.ts packages/opencode/src/session/llm-native.ts packages/opencode/test/session/llm.test.ts packages/opencode/test/session/llm-native.test.ts
• git diff --check

Stack

1. Consume native LLM events in session processing #26639
2. Add native LLM request adapter #26941
3. Compile native LLM requests in session tests #26946
4. Add native OpenAI runtime opt-in #26947 👈 current
5. Inject native LLM client into session service #27065