[lxy-9602]

Purpose

Linked issue: close #480

Blob fallback currently materializes every sequence layer before selecting the newest non-placeholder value. This adds unnecessary payload reads and collection decoding when newer layers already provide the result.

Resolve rows from newest to oldest and skip older payloads once a value, including an explicit null, has been selected. A per-file BlobReadContext bounds each read/skip run while preserving the existing field mapping, row tracking, bitmap selection, and deletion-vector processing.

Consume bounded batches directly, removing pending-batch buffering and redundant chunk offsets.

Tests

Added coverage for:

• Multi-layer fallback, explicit nulls, gaps, different batch sizes, and EOF alignment.
• Bounded reads and payload skipping with bitmap selection, deletion vectors, and row tracking.
• Covered ARRAY/MAP payloads being skipped.
• Oversized batches being rejected and temporary read requests being restored.

API and Format

Documentation

Generative AI tooling

Generated-by: OpenAI Codex (GPT-6)

[SteNicholas]

@lxy-9602, thanks for the optimization. I didn't find a correctness issue in the read/skip logic itself; the comments below are about performance in one access pattern, how the read/skip request is passed down, and layering.

[zjw1111]

+1

[SteNicholas]

LGTM.