feat(ai): add native Vercel AI Gateway language models

[rekram1-node]

Summary

• Extend the existing evaluation-only Gateway facade with native Messages, Responses, and Chat Completions APIs, reusing shared protocols.
• Default openai/gpt-* and spacexai/grok-* models to Responses, meta/muse-* to Chat, and all other model families to Messages, preserving full Gateway model IDs and explicit API selectors.
• Add typed reasoning/thinking controls, Gateway routing and upstream options, automatic caching, cache affinity, and Gateway metadata preservation.
• Preserve existing evaluation support; leave Core provider remapping, variants, and attribution integration for a later change.

Verification

• AI package: 1,755 tests passed; package typecheck and oxlint passed.
• Recorded live two-step tool loops: Claude/Gemini/DeepSeek Messages, GPT/Grok Responses, and Muse/Claude Chat.
• Repeated Claude prompt reported 4,512 cache-read tokens.
• Live invalid-key request returned non-retryable HTTP 401; fixtures also cover 403, OIDC fallback, and encrypted/signed reasoning replay.
• Checked recordings for Gateway credential leakage.

Scope

HTTP/SSE language-model support only. Gateway WebSocket, native compaction, and additional media modalities are not exposed yet. No Core wiring changes.