Pick the deepest error among the most relevant ones in each subschema

[ilia1243]

Improves best_match in the presence of anyOf / oneOf. Calculate the most relevant error in each separate subschema and choose the deepest one.

In particular, for anyOf / oneOf keywords with the only subschema, the best error is resolved as if the subschema was not enclosed by these keywords.

To reproduce:

from jsonschema import Draft202012Validator as Validator, exceptions

for applicator in "anyOf", "oneOf":
    # Should match {"properties": {"foo": {"minProperties": 2}}
    schema = {
        applicator: [
            {
                "properties": {
                    "foo": {
                        "minProperties": 2,
                        "properties": {"bar": {"type": "object"}},
                    },
                },
            },
        ],
    }
    instance = {"foo": {"bar": []}}
    error = exceptions.best_match(Validator(schema).iter_errors(instance))
    print(error)

Revert main code changes in commit b20234e preserving the tests.

Closes: #1257

---

📚 Documentation preview 📚: https://python-jsonschema--1258.org.readthedocs.build/en/1258/

[Julian]

Thanks! I'll have a look at this in the next day or two but initially looks reasonable!

[Julian]

I know I'm still behind on looking at this, but just to mention why I haven't yet merged it, I want to stare at this closer, as I'm somewhat suspicious of the types here -- specifically, it's likely true that validator_value is always a container for any builtin JSON Schema keywords, but certainly not in general (i.e. someone can invent some other one).

And similarly below I'm slightly suspicious of the schema_path business.

Not to say any of it looks wrong, but I want to convince myself if it needs any additional hardening.

And of course again thanks for the thought and PR, I'll get to it sometime soon I hope (days not weeks).

[ilia1243]

If assuming that only oneOf and anyOf can have non-empty context, then schema_path is seemingly always an integer because it was produced from enumerating https://github.com/python-jsonschema/jsonschema/blob/v4.22.0/jsonschema/_keywords.py#L340
This also implies that yes, validator_value may be something that can be enumerated, but without len.

Alternatively, assuming that best.context has errors in strict order of .schema_path (see also the same algorithms _keywords.anyOf | oneOf):

        best_in_subschemas = []
        for error in best.context:
            index = error.schema_path[0]
            if index == len(best_in_subschemas):
                best_in_subschemas.append(error)
            else:
                prev = best_in_subschemas[index]
                best_in_subschemas[index] = max(prev, error, key=key)

[ilia1243]

Superseded by #1300 taking into account #1258 (comment) and with test covering a case by @nfriedl1