Summary

• Add swarms regression scenario that reproduces Object-storage task rescheduling uses inconsistent task identity, causing task loss/misrouting after replica failure ClickHouse#1486
• Creates a Parquet file with 200 row groups, runs s3Cluster with bucket granularity, kills a replica mid-query, and verifies all rows are returned
• Validates via ProfileEvents that bucket-level splitting was active
• Add ffail to skip on non-antalya builds and versions < 26.1

Test plan

• Pre-fix build: test fails with row loss (e.g. Expected 200000 total rows, but got 165762)
• Post-fix build (PR #1493): test passes with RESULT: 200000

Made with Cursor