Add opt-in direct Gemini model support

[ahluwalij]

Summary

OpenInstinct currently treats every selected model as an AI Gateway ID. Selecting Gemini through a direct Google AI Studio key therefore needs a provider-aware path, and Gemini's tool declaration rejects the oneOf schema emitted by the existing send_message union.

This adds an opt-in Gemini 3.5 Flash-Lite (direct) workspace option. The root agent resolves that selection through the Google AI SDK; existing Gateway selections and the stored setting remain unchanged. send_message exposes a Gemini-compatible object schema while retaining the existing strict discriminated union for execution and delivery validation. The direct option is shown only when GOOGLE_GENERATIVE_AI_API_KEY is configured, and a missing key fails closed.

The browser subagent and gateway-backed web_search remain Gateway-dependent; the README documents that scope and the setup steps.

Validation

• pnpm install --frozen-lockfile
• pnpm check (83 test files, 721 tests; lint, type, formatting and repository checks passed)
• pnpm build with placeholder local DATABASE_URL, KERNEL_API_KEY, and BETTER_AUTH_URL values

The same send_message schema shape was independently verified with a delivered Linq iMessage in a downstream deployment; this PR itself has not been deployed.

[ahluwalij]

@jasonhedman could you review this direct Gemini integration, especially the send_message schema change for native Linq replies? You authored the recent Linq reply-handle and source-ownership work. The frozen install, pnpm check (721 tests), and production build pass locally.

GitHub Actions currently shows action_required for this fork PR, and the Vercel preview says Merit Systems authorization is required before deployment; neither check has run against the code yet.