[ISSUE #9728] Fix correctMinOffset handling of incomplete consume queue records

[daguimu]

Problem

In ConsumeQueue.correctMinOffset(), the size check at line 616 only guards against completely empty buffers (result.getSize() == 0). When a consume queue mapped file contains a partially written record (1-19 bytes, less than CQ_STORE_UNIT_SIZE = 20), the subsequent buffer.getLong() call throws BufferUnderflowException.

This can happen when:

1. The last mapped file in a consume queue has a partially written record due to async flush
2. CommitLog messages are lost during restart (async-flush + sync-master)
3. The broker does not truncate the incomplete record during recovery

Root Cause

The check result.getSize() == 0 does not cover the case where result.getSize() is between 1 and 19 (less than one complete CQ unit of 20 bytes). A partial record cannot be meaningfully processed.

Note: PR #10109 previously added a guard for the last mapped file check (lines 564-570), which resolved the main startup crash. This PR addresses the remaining edge case in the binary search section where a partially written record could still cause a BufferUnderflowException.

Fix

Changed the size check from:

if (result.getSize() == 0) {

to:

if (result.getSize() < ConsumeQueue.CQ_STORE_UNIT_SIZE) {

This properly handles both empty buffers and incomplete records.

Tests Added

No new tests needed — existing ConsumeQueueTest (13 tests covering correctMinOffset with various offset scenarios) all pass. The fix is a one-line defensive check improvement.

Impact

• Prevents BufferUnderflowException on partially written consume queue records
• No behavioral change for normal consume queues (all valid records are >= 20 bytes)
• Minimal, single-line change with no risk of side effects

Fixes #9728