Blob physical projection misreads empty non-null values

[zhangyang0418]

Problem

Blob columns can be misread when scanning with physical projection / binary blob handling if the column contains an empty but non-null blob value (b""). Empty blobs are encoded with position == 0 and size == 0, which can be confused with rows that have out-of-line blob data in the decode paths.

There is also a related lower-level scheduler contract issue: empty byte ranges should return an empty Bytes entry in result order, but the scheduler reassembly logic can drop empty ranges because empty intervals do not overlap any coalesced request.

Impact

After the first empty blob, subsequent blob values can be read back as empty or the scan can fail due to byte/result misalignment.

Reproduction shape

A blob column with values like:

• non-empty blob
• empty non-null blob
• non-empty blob
• null
• non-empty blob

Then scan it as binary with blob_handling="all_binary".

Expected

The scan should preserve empty non-null blobs, nulls, and all following non-empty blob payloads in order.

[zhangyang0418]

Here is a minimal Python repro:

import pyarrow as pa
import lance

val1 = b"\x01" * 1024
val2 = b"\x02" * 10240
val3 = b"\x03" * 102400
values = [val1, b"", val2, None, val3]

arr = pa.array(values, pa.large_binary())
table = pa.table(
    [arr],
    schema=pa.schema(
        [
            pa.field(
                "blobs",
                pa.large_binary(),
                metadata={"lance-encoding:blob": "true"},
            )
        ]
    ),
)

ds = lance.write_dataset(table, "/tmp/blob_empty_repro.lance")
result = ds.scanner(columns=["blobs"], blob_handling="all_binary").to_table()

assert result.column("blobs").to_pylist() == values

On affected versions, the empty non-null blob (b"") can misalign the out-of-line blob bytes, causing later non-empty values to come back empty or the scan to fail.