Lakehouse ACID with concurrent writes

A shared lakehouse table on object storage is written by multiple concurrent jobs, a streaming ingest, an hourly enrichment MERGE, a GDPR-deletion job, and a nightly compaction, all touching overlapping partitions. Object storage gives no transactions or locks. Writers are seeing lost updates, readers occasionally see partial commits, and the compaction job keeps conflicting with the MERGE. Design how ACID semantics and concurrent writes are achieved on top of a storage layer that has none of them, and how you resolve write-write conflicts.