Fix NDVariable round-trip serialization for non-float64 dtypes

[jacquelinegarrahan]

Fix NDVariable round-trip serialization for non-float64 dtypes

Branch: bugfix/ndvariable-dtype-roundtrip → main

Bug

lume/variables/ndvariable.py::coerce_default_value (a mode="before"
field_validator for default_value) rebuilt a serialized list into an array
without the model's dtype:

if isinstance(value, (list, tuple)):
    return np.asarray(value)     # dtype inferred as int64 / float64

NumPy infers int64/float64 from a plain list, but validate_default_value
(the mode="after" model validator) requires the array dtype to exactly
match self.dtype. So any NDVariable whose default_value used a different
dtype could not survive model_dump() → reconstruct, even though the
validator's own docstring promises "the model invariants are maintained."

Reproduction:

v = NDVariable(name="a", shape=(2,), dtype=np.dtype("int32"),
               default_value=np.array([1, 2], dtype="int32"))
d = v.model_dump()               # default_value -> [1, 2], dtype -> 'int32'
NDVariable(**{k: val for k, val in d.items() if k != "variable_class"})
# pydantic ValidationError: Expected dtype int32, got int64

Only the float64 default path was covered by tests
(test_round_trip_serialization), so int8/16/32, float32, etc. all failed
silently until exercised.

Fix

Read the already-validated dtype from the pydantic ValidationInfo (safe
because dtype is declared before default_value, so it is validated first)
and pass it to np.asarray, so the reconstructed array matches the declared
dtype. Falls back to NumPy's inference when no dtype is available.

Verification

pytest -q                        # 121 passed

Round-trip now succeeds for int32, float32, int8, float64, and the
no-default case:

int32   round-trip OK -> int32
float32 round-trip OK -> float32
int8    round-trip OK -> int8
float64 round-trip OK -> float64

---

🤖 Generated with Claude Code