feat: copy vision and text-only metadata from models.dev

[BrainerVirus]

Summary

• Catalog sync copies OpenCode attachment and modalities from models.dev for every matching Command Code model (including free SKUs). Unmatched models default to text-only — no vision allowlist.
• generateOpencodeModels always emits those fields so the picker knows vision vs text-only. Hy4 Preview stays text-only with a 1M context window; Gemini 3.5 Flash gets vision (and extra inputs as models.dev lists them).
• Asking the chat model “what can you do?” still will not read this spec sheet; that metadata is for OpenCode, not the LLM’s self-knowledge.

Test plan

• bun run check (148 unit tests, lint, format, tsc)
• Confirm models.json has tencent/hy4-preview with attachment: false and input ["text"]
• Confirm a vision model (e.g. google/gemini-3.5-flash) has attachment: true and "image" in modalities.input
• After npm publish, pin the new plugin version in OpenCode (not @latest while the age gate applies) and check opencode models / verbose model object for attachment and modalities

Made with Cursor

[github-actions]

🎉 This PR is included in version 0.7.0 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀