After alot of debugging and retrying i got to build a host correctly with image through the api with autostart of the virtual machine in libvirt.
If you touch compute_attributes you have to complete the whole block with more data to get it done correctly.
Attaching example for other users having issues with this:

curl -s -X POST -H "Accept:application/json" -H "Content-Type:application/json" \
     -k -u $USER:$PASS \
-d  \
'{"host":{  
      "name":"'$NAME'",
      "hostgroup_id":"1",
      "compute_resource_id":"1",
      "compute_profile_id":"1",
      "environment_id":"1",
      "puppet_ca_proxy_id":"",
      "puppet_proxy_id":"",
      "salt_environment_id":"3",
      "salt_proxy_id":"3",
      "salt_module_ids":[  
         "" 
      ],
      "puppetclass_ids":[  
         "" 
      ],
      "managed":"true",
      "progress_report_id":"[FILTERED]",
      "type":"Host::Managed",
      "compute_attributes":{  
         "cpus":"1",
         "memory":"805306368",
         "start":"1",
            "volumes_attributes": {
                "0": {
                    "_delete": "",
                    "pool_name": "default",
                    "capacity": "10G",
                    "allocation": "0G",
                    "format_type": "qcow2" 
                },
                "new_volumes": {
                    "_delete": "",
                    "pool_name": "default",
                    "capacity": "10G",
                    "allocation": "0G",
                    "format_type": "raw" 
                }
            },
         "image_id":"/var/lib/libvirt/images/foo.img" 
      },
      "architecture_id":"1",
      "subnet_id":"1",
      "operatingsystem_id":"1",
      "provision_method":"image",
      "build":"1",
      "medium_id":"2",
      "ptable_id":"9",
      "disk":"",
      "enabled":"1",
      "model_id":"",
      "comment":"",
      "overwrite":"false" 
     }}' \
     $FOREMAN_URL/api/v2/hosts | json_pp

I have this exact same problem. It would be great if we could use "start":"1" without explicitly setting the rest of the compute_attributes. These should already be set through the compute profile and shouldn't have to be set again when posting to the api.