fix: refine Google Gemini parameter applicability

[guillaumegay13]

Summary

• remove Google Gemini presence/frequency penalty params because live generateContent rejects them for all four cataloged Gemini models
• relax generationConfig.topK lower bound from 1 to 0, matching live validation (topK: 0 accepted, topK: -1 rejected as nonnegative)

Verification

• Official docs checked today:
  • https://ai.google.dev/api/generate-content
  • https://ai.google.dev/gemini-api/docs/thinking
• Live Google API matrix with key from .env.keys:
  • getModel confirmed gemini-3.5-flash exists today as 3.5-flash-05-2026
  • sampling params without penalties are accepted with thinking enabled on gemini-3.5-flash, gemini-2.5-pro, gemini-2.5-flash, and gemini-2.5-flash-lite
  • presencePenalty: 0.1 and frequencyPenalty: 0.1 are rejected on all four with Penalty is not enabled
  • all listed post-patch params were sent together per YAML row and returned HTTP 200 for all four models
• npm run validate
• npm test
• npm run typecheck
• npm run lint
• npm run build
• npm run guard:params
• npx prettier --check .
• git diff --check

[vercel]

@guillaumegay13 is attempting to deploy a commit to the Manifest Team on Vercel.

A member of the Team first needs to authorize it.