in https://schema.org/mainEntityOfPage we have

{
    "@context": "https://schema.org",
      "@type": "Restaurant",
      "mainEntityOfPage": {
         "@type": "WebPage",
         "@id": "http://cathscafe.example.com/"
      },
      "name": "Cath's Cafe",
      "openingHours": "Mo,Tu,We,Th,Fr,Sa,Su 11:00-20:00",
      "telephone": "+155501003344",
      "hasMenu": "/menu"
}

This is verbose, and boring, and parses to a different structure than the RDFa and Microdata samples nearby.

It is rarely useful to go to the trouble of saying that a web page is a WebPage. We have done it here because having

 "mainEntityOfPage": { "@id": "http://cathscafe.example.com/" }

... on its own breaks the informal schema.org convention of supplying types where there is a syntactic context to include them easily.

However, schema.org's JSON-LD context file declares mainEntityOfPage to default to identifiers, so we could just write:

 "mainEntityOfPage": "http://cathscafe.example.com/" 

... and have this parse out to a URI. Let's do that.

The verbose / boring form is not wrong just a little bit pointless and slightly unusual. Having our examples across syntaxes encode exactly the same graph is generally a good thing and a sensible default. There may be cornercases (e.g. markup in literals, which microdata can't do; also breadcrumbs) where examples in different syntaxes can diverge, but let's try to minimize this practice.