to_bytes silently rescales a Decimal with a negative scale

pyiceberg.conversions.to_bytes のエントリポイントと decimal_to_unscaled から始め、続いて DecimalType(10, 2) 用に提供された Decimal の再現コードを実行します。完了条件は、一致しない符号付きスケールが再スケーリングされずに拒否される一方で、型のスケールにすでに一致している値は引き続き正しくラウンドトリップすることです。pyiceberg/manifest.py と pyiceberg/io/pyarrow.py への影響を確認してください。

to_bytes for DecimalType takes the absolute value of the exponent before comparing it
against the type's scale:

_, digits, exponent = value.as_tuple()
exponent = abs(int(exponent))
if exponent != primitive_type.scale:
    raise ValueError(...)

A Decimal carries the negated scale as its exponent, so a value with a negative scale
passes this check as if it had the matching positive one. decimal_to_unscaled then uses
only the digits, and the exponent is dropped:

sign, digits, _ = value.as_tuple()
return int(Decimal((sign, digits, 0)).to_integral_value())

The value is written four orders of magnitude off, with no error.

Reproduction

from decimal import Decimal
from pyiceberg.conversions import to_bytes, from_bytes
from pyiceberg.types import DecimalType

t = DecimalType(10, 2)
print(from_bytes(t, to_bytes(t, Decimal("1E+2"))))   # 0.01, expected 100.00
print(from_bytes(t, to_bytes(t, Decimal("5E+1"))))   # 0.50 for decimal(10, 1) -> 0.5, expected 50.0

This is not an exotic input. Decimal.normalize() produces exactly this form:

Decimal("100").normalize()   # Decimal('1E+2')

so a value that has been normalized, or that comes out of arithmetic that trims trailing
zeros, hits it.

Impact

to_bytes writes the lower_bounds and upper_bounds of a data file
(pyiceberg/manifest.py, _write_data_file_statistics), and the same conversion is used
in pyiceberg/io/pyarrow.py. A bound written as 0.01 instead of 100.00 makes scan
planning prune files that do hold matching rows, so a query silently returns fewer rows
than it should.

Values with a positive scale are unaffected: Decimal("100.00") round-trips correctly.

Suggested fix

Compare the signed scale, -exponent, against primitive_type.scale, so a mismatching
value is rejected instead of being silently rescaled. Rescaling the value to the type's
scale would be the other option, but that widens the contract of a function that today
requires an exact match.

Happy to open a PR.