Validation allows more type promotion than the evaluator

[nsmith-]

For example, the correction

{
  "schema_version": 2,
  "corrections": [
    {
      "name": "hi",
      "version": 2,
      "output": { "name": "hi", "type": "real" },
      "inputs": [{ "name": "blah", "type": "int" }],
      "data": {
        "nodetype": "category",
        "input": "blah",
        "content": [{ "key": "a", "value": "3.0" }]
      }
    }
  ]
}

has a content that should be a float but pydantic will happily validate and type-promote a string, e.g.

parsed = cs.CorrectionSet.model_validate_json(data)
print(parsed.corrections[0].data)

produces

nodetype='category' input='blah' content=[CategoryItem(key='a', value=3.0)] default=None

A similar issue is the source of the TODO in

correctionlib/tests/test_core_valid.py

Lines 47 to 48 in b6a3485

	# TODO: this is not detected by the pydantic model yet
	hard_json = '{"schema_version":2, "corrections": [{"name": "hi", "version": 2, "output": {"name": "hi","type": "real"}, "inputs": [{"name":"blah", "type": "int"}], "data": {"nodetype": "category", "input": "blah", "content": [{"key": "a", "value": 3.0}]}}]}'

A solution could either:

• Do the same type promotion inside the C++ evaluator, or
• Configure pydantic to treat such cases as errors with strict mode

A potential workaround (or at least mechanism to catch such cases) in the correction validate CLI is to also construct an evaluator from the parsed model.