feat(mitm): add Antigravity reasoning effort overrides

[trfi]

Summary

Add per-source-model reasoning effort overrides to the Antigravity MITM Proxy alongside the existing destination model mappings.

Each Antigravity model row now supports:

• Default
• None
• Minimal
• Low
• Medium
• High
• XHigh
• Max

Default preserves the thinking configuration sent by Antigravity. Explicit selections override the request and are translated through 9Router's existing unified thinking pipeline.

Changes

• Add an accessible reasoning selector beside each Antigravity model mapping.
• Allow model and reasoning overrides to be configured independently.
• Support reasoning-only overrides without changing the source model.
• Store structured { model, reasoningEffort } entries.
• Preserve compatibility with legacy string-based model mappings.
• Normalize cached aliases for the standalone MITM runtime.
• Validate reasoning values at the API boundary.
• Remove conflicting Gemini thinking fields when an explicit effort is selected.
• Keep native Antigravity thinking fields unchanged when effort is Default.
• Add focused tests for mapping normalization and request transformation.
• Document persistence, reasoning values, and safety behavior.

Backward compatibility

Existing mappings remain valid:

{
  "gemini-3-flash-agent": "provider/model-id"
}

They are normalized at the API and standalone MITM boundaries into:

{
  "gemini-3-flash-agent": {
    "model": "provider/model-id"
  }
}

No database migration is required.

Verification

• Focused ESLint check passed.
• CommonJS MITM syntax checks passed.
• Mapping and request-transformation smoke assertions passed.
• Next.js development server started successfully.
• Mapping API route compiled successfully.
• git diff --check passed.

Notes

• The repository's root dependencies do not include Vitest, so the focused Vitest file was not executed in the existing workspace.
• Equivalent core normalization and request-transformation assertions were executed directly with Node.
• Antigravity tab-completion models remain excluded from MITM rerouting.