id:int score:int tag:string
primary key: id
no partition key
bucket count: 1

Generated by Apache Paimon Java release-2.0.0.
file format: orc
deletion-vectors.enabled: true
deletion-vectors.merge-on-read: false
write-only: true (explicit fixture commits produce snapshots 2, 4, 5)
primary-key btree index: score
target-file-size: 8 kb
compaction.force-rewrite-all-files: true

Snapshot semantics:
snapshot-1 APPEND: ids 1..2000; score=id%10; even tag=keep, odd tag=drop.
snapshot-2 COMPACT: full compaction builds source-backed score BTree payloads; source-file counts per payload are [2]; 2000 rows.
snapshot-3 APPEND: insert (2001,0,keep) and (2002,5,late_drop); indexed compacted files and a visible unindexed APPEND-source file coexist; 2002 rows.
snapshot-4 COMPACT: delete id=10 and update id=20 to (77,updated), then lookup compact in one fixture commit; deletion vectors coexist with indexed files; 2001 rows.
snapshot-5 COMPACT: full compaction rebuilds the BTree source groups; 2001 rows.

Fixture construction note: with deletion vectors enabled and merge-on-read disabled, Java and C++ batch scans skip literal level-0 files. Snapshot 3 therefore promotes Java-prepared files to the maximum manifest level while retaining FileSource.APPEND. The source-backed PK BTree policy indexes only COMPACT files above level 0, making these files visible but deliberately unindexed. Snapshot 4 publishes the Java-prepared compact outputs and index changes while omitting their transient, never-active append inputs. These are test-fixture constructions, not ordinary end-user write workflows.

At snapshots 4 and 5, predicate score=0 AND id<=1500 returns ids 30,40,...,1500; id=10 is deleted and id=20 has score 77.
