feat(session/llm): scaffold OpenAI Realtime (WebSocket) transport

[WonderJL]

Summary

• Scaffolds an OpenAI Realtime API (WebSocket) transport for the session LLM stream.
• Ships behind a new experimental.openai_realtime flag (default false) — no behavior change until the flag is enabled and the wiring lands.
• Includes a pure event mapper (Realtime JSON → LLMEvent) with unit tests, a stub stream() that fails loudly, and a design note documenting the protocol mapping and open questions.

Why a scaffold, not a full transport?

The Realtime event surface is wide and additive; the right dispatch shape inside session/llm.ts is a maintainer call. Landing the scaffold first lets reviewers shape that integration without a churny diff. This PR is intentionally inert at runtime — stream() throws RealtimeNotImplemented if accidentally invoked.

What changes

• packages/opencode/src/session/llm/realtime.ts (new) — adapterState(), toLLMEvents(), REALTIME_URL, RealtimeNotImplemented, stub stream(). The mapper shape mirrors ai-sdk.ts's adapterState() / toLLMEvents() on purpose so the eventual dispatch site treats both adapters uniformly.
• packages/opencode/src/session/llm/REALTIME-DESIGN.md (new) — scope, non-goals, protocol mapping table, lifecycle plan, auth, open questions for maintainers.
• packages/opencode/src/config/config.ts — adds experimental.openai_realtime: Schema.optional(Schema.Boolean).
• packages/opencode/test/session/llm-realtime.test.ts (new) — four unit tests covering text turns, function-call argument streaming, unknown-event passthrough, and failed-response status.

Realtime → LLMEvent mapping (summary)

Realtime event	LLMEvent
response.created	stepStart
response.output_item.added (message)	textStart
response.output_text.delta	textDelta
response.output_text.done	textEnd
response.output_item.added (function_call)	toolInputStart
response.function_call_arguments.delta	toolInputDelta
response.function_call_arguments.done	toolInputEnd
response.output_item.done (function_call)	toolCall (with parsed arguments)
response.done	stepFinish + finish

Unknown event types (audio., rate_limits., etc.) drop to [] rather than fail.

Open questions (called out in the design note)

1. Register Realtime as a separate providerID (openai-realtime) or keep it under openai and switch transport by flag?
2. Reuse ProviderTransform.tools for tool-schema translation, or build a Realtime-specific path?
3. How to treat gpt-5* reasoning items (which aren't first-class in Realtime yet) — block at status() or pass through?

Test plan

• bun run typecheck (packages/opencode) — clean.
• bun test test/session/llm-realtime.test.ts — 4/4 pass.
• Once wiring lands in a follow-up: live smoke against gpt-4o-realtime-preview / gpt-5.x-realtime with a tool-call prompt.

Notes for reviewers

• The scaffold deliberately stops short of opening a socket — happy to push the lifecycle (acquireRelease + AbortSignal bridge + idle timeout) in a follow-up once the integration shape is agreed.
• This is independent of fix(session): add idle timeout to LLM streaming response #29348 (idle-timeout PR) but the eventual live wiring will reuse the same Stream.timeoutOrElse idle guard so both transports fail-fast on stalls.

[ghost]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[ghost]

Hey! Your PR title feat(session/llm): scaffold OpenAI Realtime (WebSocket) transport doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.