fix(aggregate): preserve listagg result ownership

[ChaomingZhangCN]

Purpose

Linked issue: N/A

FieldListaggAgg returned a std::string_view backed by a mutable member string shared by all calls to the same aggregator instance.

The merge path reuses one field aggregator across primary keys, while SortMergeReaderBatchProducer may retain results for multiple keys before converting the batch to Arrow. As a result, aggregating a later key could overwrite or invalidate the string data referenced by an earlier key, causing cross-key data corruption.

This change:

• builds each concatenated result independently;
• stores the result in a pool-backed BinaryString;
• removes the shared mutable result buffer;
• preserves the lifetime of previously returned aggregation results.

Tests

Added:

• FieldListaggAggTest.TestResultOwnershipAcrossAggregations for the aggregator result-lifetime contract;
• WriteAndReadInteTest.TestPKListAggPreservesResultsAcrossKeys for the real PK merge path in write_and_read_inte_test.cpp.

The integration test creates a PK aggregation table, writes two keys over two commits, and verifies that each key retains its independently accumulated listagg value.

Before the production fix, the integration test fails deterministically for all three enabled formats (Parquet, ORC, and Avro):

Expected: alpha,beta
Actual:   one,two\0\0\0

After the fix:

• paimon-write-and-read-inte-test builds successfully;
• TestPKListAggPreservesResultsAcrossKeys: 3/3 format variants passed;
• neighboring TestPKSimple: 3/3 format variants passed;
• paimon-core-test builds and links successfully on macOS arm64 after replacing the non-portable BIGINT test literals;
• clang-format, cpplint, codespell, and git diff --check passed.

The focused FieldListaggAggTest.* execution is still blocked in this local build by an existing macOS static/shared plugin registration issue (Could not find a FileFormatFactory implementation class for format 'parquet'). This does not affect the integration-test target above.

API and Format

No public API, storage format, or protocol changes.

Documentation

No documentation changes are required.

Generative AI tooling

Generated-by: OpenAI Codex (GPT-5)

[lxy-9602]

+1