[Feature] Add gemini-3.7-flash with thinking level mapping and update google defaults

[Kenmege]

Problem

gemini-3.7-flash was present in upstream OpenRouter catalogs, but under the direct google and google-vertex providers in Prime Agent, its reasoning/thinking level map was unmapped ({"off": null}), preventing /effort high or high-reasoning level execution from mapping properly to Google GenAI's HIGH thinking level. In addition, the default model per provider for google and google-vertex was pointing to older preview models (gemini-3.1-pro-preview).

Evidence

• Google GenAI API and Vertex AI APIs support thinkingLevel: "MINIMAL", "LOW", "MEDIUM", "HIGH" for gemini-3.7-flash.
• OpenRouter lists google/gemini-3.7-flash with reasoning supported and thinking levels low/medium/high.
• models.dev lists gemini-3.7-flash for google with reasoning enabled.

Proposed change

1. Update packages/ai/scripts/generate-models.ts to map thinking levels for Gemini 3.x Flash models (gemini-3.7-flash, etc.) to {"off": null, "minimal": "MINIMAL", "low": "LOW", "medium": "MEDIUM", "high": "HIGH"}.
2. Add explicit gemini-3.7-flash entries to google-vertex provider and fallback generator registration for google.
3. Update defaultModelPerProvider for google and google-vertex in packages/coding-agent/src/core/model-resolver.ts to gemini-3.7-flash.
4. Add unit test suite in packages/ai/test/google-gemini-37-flash.test.ts.

Ready branch

Validated on Kenmege/prime-agent branch feat/gemini-3.7-flash (branched from current main):

• npm run check clean from repo root (biome 949 files, tsgo --noEmit, installer render, browser smoke).
• packages/ai/test/google-gemini-37-flash.test.ts passed (2/2 tests).
• npm run build executed cleanly.

The branch is push-button ready at https://github.com/Kenmege/prime-agent/tree/feat/gemini-3.7-flash.