What is the easiest way to do streaming parsing of non-uniform JSON data using System.Text.Json?

I am trying to stream-parse JSON in the SPARQL JSON results format, e.g.:

{ "head": {
    "vars": [ "s" , "p" , "o" ]
  } ,
  "results": {
    "bindings": [
      { 
        "s": { "type": "uri" , "value": "https://example.com/ontology/al/bnode/1" } ,
        "p": { "type": "uri" , "value": "http://www.w3.org/1999/02/22-rdf-syntax-ns#type" } ,
        "o": { "type": "uri" , "value": "https://example.com/ontology/al/Module" }
      } ,
      { 
        "s": { "type": "bnode" , "value": "b0" } ,
        "p": { "type": "uri" , "value": "http://www.w3.org/1999/02/22-rdf-syntax-ns#rest" } ,
        "o": { "type": "uri" , "value": "http://www.w3.org/1999/02/22-rdf-syntax-ns#nil" }
      } ,
      { 
        "s": {
          "type": "triple" , 
          "value": {
            "subject":  { "type": "uri" , "value": "https://example.com/ontology/al/bnode/1798" } ,
            "predicate": { "type": "uri" , "value": "https://example.com/ontology/al/type" } ,
            "object":   { "type": "uri" , "value": "https://example.com/ontology/al/bnode/6" }
          }
        } ,
        "p": { "type": "uri" , "value": "https://example.com/ontology/al/dimensions" } ,
        "o": { "type": "bnode" , "value": "b0" }
      } ,
      { 
        "s": { "type": "uri" , "value": "https://example.com/ontology/al/bnode/1" } ,
        "p": { "type": "uri" , "value": "https://example.com/ontology/al/publisher" } ,
        "o": { "type": "literal" , "value": "Example" }
      }
  ]
 }
}

The head/vars mentions variables, which are specific to the query at hand and bound to specific values in the results/bindings array.

Normally I would JsonSerializer.DeserializeAsyncEnumerable<T> but since the top-level element is not an array, that does not seem workable. That the individual variable bindings are non-uniform is not a problem per se, as T could be the JsonElement type.

The data is comming from an HTTP stream. For a quasi-long-running query, it would give a better UX if the results could be incrementally integrated into the UI of the client application.