fix premature model construction

[d-v-b]

2 fixes in this PR:

• removes the default attributes for v3 arrayspec
• allows from_zarr to work by constructing the final group from dicts instead of arrayspec / groupspec instances

[dstansby]

Should this not be dict[str, GroupSpec | ArraySpec], and pydantic will handle the casting if it's given dict[str, object] as input?

[d-v-b]

we are filling members with dicts produced by ArraySpec.model_dump() / GroupSpec.model_dump(). If we don't do that, then pydantic will reject models that declared specialized members, because pydantic uses nominal typing (isinstance checks) when the input is a basemodel

[dstansby]

Won't the specialised members be sub-classes of ArraySpec or GroupSpec though, and therefore would pass the isinstance check?

[d-v-b]

this is what will happen:

from pydantic_zarr.experimental.v3 import ArraySpec, GroupSpec
import numpy as np

class MyArray(ArraySpec):
    ...

class MyGroup(GroupSpec):
    members: dict[str, MyArray]

MyGroup(attributes={},members={'a': ArraySpec.from_array(np.arange(10))})
"""
pydantic_core._pydantic_core.ValidationError: 1 validation error for MyGroup
members.a
  Input should be a valid dictionary or instance of MyArray [type=model_type, input_value=ArraySpec(zarr_format=3, ...), dimension_names=None), input_type=ArraySpec]
    For further information visit https://errors.pydantic.dev/2.11/v/model_type
"""

[d-v-b]

isinstance(ArraySpec(...), SpecializedArraySpec) will fail, which is why we need to turn the arrayspec / groupspec members into dicts

[dstansby]

Okay - sorry, this is somewhat off-topic on this PR

[d-v-b]

no worries!