feat(server): list experimental voice models - #57194

Why

iOS and desktop clients need to discover which voice models the server can start (and which transports each supports) before initiating a call, rather than guessing a model and hitting a 503 Service Unavailable.

What

• Extend Realtime.engines (packages/ai/src/experimental/realtime.ts) with each engine's default model (gpt-live-1, gpt-realtime-2.1, grok-voice-3.0, gemini-3.8-live) and supported transports, and validate connect against resolved.transports so the listing and session handshake cannot drift.
• Add Voice.Service.models() (packages/core/src/voice.ts), sharing findConnection(provider) with Voice.Service.session(input) to check whether an API key or non-empty environment variable connection exists without resolving secret material or calling providers.
• Expose GET /api/experimental/voice/models (experimental.voice.models) in @opencode/protocol, @opencode/server, @opencode/client, and OpenAPI specs.
• Known limitation: Integration tracks one active connection per provider; voice picks the first stored API key or non-empty environment variable connection and skips OAuth/ChatGPT sign-ins.

Call-site example

const { models, default: defaultModel } = await client.voice.models()
// {
//   models: [
//     { model: "gpt-live-1", provider: "openai", transports: ["webrtc"], available: true },
//     { model: "gpt-realtime-2.1", provider: "openai", transports: ["webrtc", "websocket"], available: true },
//     { model: "grok-voice-3.0", provider: "xai", transports: ["websocket"], available: false },
//     { model: "gemini-3.8-live", provider: "google", transports: ["websocket"], available: false },
//   ],
//   default: "gpt-live-1",
// }

Verification

• packages/ai: bun typecheck, bun test test/realtime.test.ts
• packages/core: bun typecheck, bun run script/test.ts test/voice.test.ts
• packages/server: bun typecheck, bun test test/voice.test.ts
• packages/client: bun typecheck, bun test test/promise.test.ts, bun run check:generated
• packages/protocol & services/www: bun run check:generated

Added by the opencode bot (AI-generated).

Public API: adds GET /api/experimental/voice/models and client.voice.models() (backwards compatible)