[piyush15102003]

Summary

Reading a Native result that spans more than one block lost two rows at every block boundary: the last row of a block and the first row of the next were returned as all-null rows, with the row count unchanged and no error.

AbstractBinaryFormatReader reads one row ahead. When the read-ahead starts a new block, NativeFormatReader.readBlock() calls setSchema(), which replaced both row buffers. The read-ahead then wrote the new block's first row into the old nextRecord array, and next() wrapped the freshly allocated, empty currentRecord.

Changes:

• setSchema() allocates currentRecord / nextRecord only when they are null or the column count differs, so the first call (constructor, RowBinaryWithNamesAndTypes header) still allocates them.
• next() takes the record reference before the read-ahead, so the returned row does not depend on what the read-ahead does to the fields.

Closes #3151

Checklist

• Closes [client-v2] Native format: two rows are silently lost and returned as all-null at every block boundary #3151
• Unit tests covering the common scenarios were added (NativeFormatReaderBlockBoundaryTest: 25 rows in 10-row blocks via next() and via getters, and 1-row blocks). They check every value, not only the row count, and all three fail on main (Lists differ at element [9]: 9 != null).
• A human-readable description of the changes was provided to include in CHANGELOG (history/latest/3151.md)

---

Note

Medium Risk
Touches core Native/binary row iteration where incorrect behavior silently corrupted values; the fix is narrow and covered by new regression tests.

Overview
Fixes silent data loss when reading multi-block Native query results in client-v2: at every block boundary the last row of one block and the first row of the next were replaced by all-null rows while the row count stayed the same.

The reader read-aheads one row; starting the next block re-publishes the schema via setSchema(), which used to always allocate new currentRecord / nextRecord buffers and left next() wrapping empty arrays after the read-ahead. setSchema() now reuses those buffers when the column count is unchanged, and next() captures the row array reference before readNextRecord() so returned rows are not tied to buffer swaps during read-ahead.

Adds NativeFormatReaderBlockBoundaryTest (multi-block and 1-row blocks, via next() and typed getters) and a changelog note for #3151.

^{Reviewed by Cursor Bugbot for commit ed0e317. Bugbot is set up for automated code reviews on this repo. Configure here.}