fix(mcp): handle non-string metadata in cloud sync handler

[DelphiaDan]

Summary

• The cloud sync handler queries a collection for its first document's metadata to extract the codebasePath
• Depending on the vector DB backend, metadata may already be a parsed JS object rather than a JSON string
• JSON.parse() is called unconditionally on line 86, which throws when the value is not a string
• The catch block logs a warning and the codebase path never registers, so search_code reports "codebase not indexed" even after successful indexing

Fix

Add a type guard to only parse when metadata is actually a string:

const metadata = typeof metadataStr === 'string' ? JSON.parse(metadataStr) : metadataStr;

Fixes #226