While nested documents provide additional features, we don't needed nested documents for everything. Predicates on a single field when we're not using the nested predicate, in particular, work perfectly well on flattened object fields, and actually perform better since we don't need a runtime join.
To get the best of both worlds, when one can spare the extra disk space it requires, we could allow object fields to be stored both as a nested document and flattened in their parent. Elasticsearch allows that with the attribute include_in_parent (see https://www.elastic.co/guide/en/elasticsearch/reference/current/nested.html#nested-params).
API-wise, this could just be an additional value for the ObjectStructure enum: ObjectStructure.NESTED_AND_FLATTENED. We could then take advantage of this internally: when nesting is not necessary (no nested predicate, no filter in sorts and aggregations), we can just use the flattened field.