[BugFix] Don't advance batch-apply rssid offset for zero-segment op_writes

Row-mode partial update on lake PK produced duplicate PK rows: in the
batch-apply path, MetaFileBuilder::add_rowset advanced
_pending_rowset_data.assigned_segment_idx via get_rowset_id_step() on every
op_write, including a leading ZERO-segment op_write (get_rowset_id_step
returns 1 for an empty rowset). Because the pending rowset still had
segment_metas_size()==0, the next segment-bearing op_write re-entered the
first-call branch and stamped its first segment with a non-positional
segment_idx (assigned_segment_idx + 0 = 1). At read the segment's effective
rssid (rowset.id() + segment_idx) landed one slot above where the base-row
delete vector was keyed at apply, so the superseded base rows were never
hidden -> duplicate PKs. Only advance assigned_segment_idx for op_writes
that actually deposit segments.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: meegoo

be/src/storage/lake/meta_file.cpp

@@ -1498,7 +1498,17 @@ void MetaFileBuilder::add_rowset(const RowsetMetadataPB& rowset_pb, const std::m
     _pending_rowset_data.dels.insert(_pending_rowset_data.dels.end(), dels.begin(), dels.end());
 
     // Track cumulative rssid slots already assigned when batch applying multiple opwrites.
-    _pending_rowset_data.assigned_segment_idx += get_rowset_id_step(rowset_pb);
+    // FIX (row-mode partial-update duplicate PK regression): only advance the rssid offset for a
+    // contribution that actually deposited segments. get_rowset_id_step() returns 1 even for a
+    // 0-segment op_write, so a leading zero-segment op_write would bump assigned_segment_idx while
+    // leaving the pending rowset at segment_metas_size()==0; the next segment-bearing op_write then
+    // re-enters the first-call branch above and stamps its first segment with
+    // segment_idx = assigned_segment_idx(1) + 0 = 1 (non-positional). At read the segment's effective
+    // rssid = rowset.id() + segment_idx lands one slot above where the base-row delete vector was keyed
+    // at apply (the positional rssid), so the superseded base rows are never hidden -> duplicate PKs.
+    if (rowset_pb.segment_metas_size() > 0) {
+        _pending_rowset_data.assigned_segment_idx += get_rowset_id_step(rowset_pb);
+    }
 }
 
 Status MetaFileBuilder::set_final_rowset() {