fix(ai): update stale catalog test models

[quick-ricon]

Summary

• update the Copilot overflow test to a current Gemini model
• move Groq's Qwen reasoning-level override to the current catalog model
• update the matching Groq regression fixture

This repairs the two stale model IDs that currently stop main during TypeScript validation.

Validation

• npm run check
• focused Groq reasoning regression: 1 passed
• current model-data hydration confirms both replacement IDs and the expected Qwen mapping
• git diff --check

The broad local ./test.sh path was not run from this credentialed session because the KKL runtime-release runbook delegates that interruption-sensitive gate to hosted CI.