Blob encoding metadata not applied when writing Lance dataset via Ray Data mapfunction

Steps to reproduce:

import lance
import pyarrow as pa
import ray.data
import lance
import pyarrow as pa
import pandas as pd
from pathlib import Path
import random

schema = pa.schema([
    pa.field("id", pa.int64()),
    pa.field("video_name", pa.string()),
    pa.field("video_size", pa.int64()),
    pa.field(
        "video_data",
        pa.large_binary(),
        metadata={"lance-encoding:blob": "true"}
    ),
])

mp4_files_dir="/DATA/workshop/personal/Lance/HDFS"
mp4_dir = Path(mp4_files_dir)
mp4_files = list(mp4_dir.glob("*.mp4"))

if not mp4_files:
    print(f"在目录 {mp4_files_dir} 中未找到 MP4 文件")

print(f"找到 {len(mp4_files)} 个 MP4 文件")

class Video:
    def __init__(self):
        pass

    def __call__(self, row):
        print(f"this row {row}")
        mp4_file = random.choice(mp4_files)
        try:
            with open(mp4_file, 'rb') as f:
                video_bytes = f.read()

            row[f"video_name"] = mp4_file.name
            row[f"video_size"] = mp4_file.stat().st_size
            row[f"video_data"] = video_bytes

            print(f"已加载: {mp4_file.name} ({len(video_bytes)} 字节)")

        except Exception as e:
            print(f"读取文件 {mp4_file} 时出错: {e}")
        return row


LANCE_PATH = "s3://tmp/test1.lance"


def write():

    ray.data.range(100).map(Video).write_lance(
        LANCE_PATH, mode="create", min_rows_per_file=100, storage_options=storage_options, schema=schema)


def read():
    df = lance.dataset(LANCE_PATH, storage_options=storage_options)
    print(
        df.to_table().to_pandas()
    )

print(ray.data.range(100).map(Video).schema())
write()

Expected behavior:

The video_datafield should be stored with blob encoding as specified in the schema metadata.

Actual behavior:

The field is stored as regular binary data without blob encoding.

Show table result

Start by running the provided Ray Data reproduction and compare the schema returned by ray.data.range(100).map(Video).schema() with the schema passed to write_lance. Trace how the video_data field metadata is propagated through write_lance and into lance.dataset, then verify that the stored field retains blob encoding metadata.