Viewless timeseries collections eliminate the separate view and buckets namespace structure, but many of the failing operations on sharded collections (particularly upserts, findAndModify, and multi-document transactions) still contain code paths that attempt to call getTimeseriesViewNamespace() on what is now a regular collection namespace.

On unsharded collections these operations execute fine, but in these paths we're doing an invalid namespace translation to the legacy timeseries namespaces - leading to invariant failures such as this one in testing.