Trouble using Profile / Structure Definition for resource validation not working as expected

Trying to custom-validate a patient resource.
For testing proof of concept, I added a non-existent field level.

This is my Structure Definition which outputs a valid result:

"snapshot": [
    ...
    {
    "id": "Patient.level",
    "path": "Patient.level",
    "short": "easy | hard | expert | hardcore",
    "definition": "patient level",
    "comment": "comment",
    "requirements": "Needed for testing",
    "min": 1,
    "constraint": [
      {
        "key": "lev-1",
        "severity": "error",
        "human": "bla bla bla",
        "expression": "$this.exists()"
      }
    ],
    "condition": [
      "lev-1"
    ],
    "type": [
      {
        "code": "code"
      }
    ],
    "isSummary": true
  }

This results in a valid resource.
But, if I only add the line:

"min": 1,

The result of the validation is:

Found issue:ERROR - Bundle.entry.resource - Profile http://hl7.org/fhir/StructureDefinition/Patient, Element 'Bundle.entry.resource.level': minimum required = 1, but only found 0

I cannot seem to change the severity to Warning, no matter what.

That’s because you can’t do what you are trying to do - profiles cannot add new elements that don’t already exist

I see. I was trying to isolate the problem so I added a new element, but If I try to do the exact same thing to an existing element - I have the same issue.

what’s an example of that then?

same thing but Patient.gender instead of Patient.level:

{
    "id": "Patient.gender",
    "path": "Patient.gender",
    "short": "male | female | other | unknown",
    "definition": "Administrative Gender - the gender that the patient is considered to have for administration and record keeping purposes.",
    "comment": "The gender may not match the biological sex as determined by genetics, or the individual's preferred identification. Note that for both humans and particularly animals, there are other legitimate possibilities than M and F, though the vast majority of systems and contexts only support M and F.  Systems providing decision support or enforcing business rules should ideally do this on the basis of Observations dealing with the specific gender aspect of interest (anatomical, chromosonal, social, etc.)  However, because these observations are infrequently recorded, defaulting to the administrative gender is common practice.  Where such defaulting occurs, rule enforcement should allow for the variation between administrative and biological, chromosonal and other gender aspects.  For example, an alert about a hysterectomy on a male should be handled as a warning or overrideable error, not a \"hard\" error.",
    "requirements": "Needed for identification of the individual, in combination with (at least) name and birth date. Gender of individual drives many clinical processes.",
    "min": 0,
    "constraint": [
      {
        "key": "gen-7",
        "severity": "error",
        "human": "bla bla bla",
        "expression": "$this.exists()",
        "source": "http://hl7.org/fhir/StructureDefinition/Element"
      }
    ],
    "condition": [
      "gen-7"
    ],
    "type": [
      {
        "code": "code"
      }
    ],
    "isSummary": true
  },