[Python][Parquet] read_schema drops extension types (UUID returned as fixed_size_binary[16])

[Kuinox]

Summary

UUID extension types are preserved in tables but dropped by pyarrow.parquet.read_schema, creating an asymmetry between the table’s schema and the schema read from Parquet metadata.

Steps to Reproduce

import pyarrow as pa
import pyarrow.parquet as pq
from pathlib import Path
import tempfile

data = [
 b'\xe4`\xf9p\x83QGN\xac\x7f\xa4g>\x4b\xa8\xcb',
 b'\x1et\x14\x95\xee\xd5C\xea\x9b\xd7s\xdc\x91BK\xaf',
 None,
]
table = pa.table([pa.array(data, type=pa.uuid())], names=["ext"])
print("table schema type:", table.schema.field("ext").type) # extension<arrow.uuid>

path = Path(tempfile.gettempdir()) / "uuid_ext_test.parquet"
pq.write_table(table, path, store_schema=False)

print("read_schema type:", pq.read_schema(path).field("ext").type)
print("read_table schema type:", pq.read_table(path).schema.field("ext").type)

Expected Behavior

read_schema(path) should yield the same type as the table schema (and read_table), i.e., extension<arrow.uuid>.

Actual Behavior

read_schema(path) returns fixed_size_binary[16], while the original table.schema and read_table(path).schema both report extension<arrow.uuid>, so metadata-based schema inspection drops the extension type.

Notes

• Observed with the current pyarrow wheel (22.0.0) and current main sources.
• ParquetFile(...).schema_arrow preserves the extension type; read_schema does not

Component(s)

Python, Parquet

[rohanjain101]

Similar to #48169