When the user doesn't specify "on" field for $merge, it default to _id (+ shard key for sharded collections).

When the default value is used for "on", there is no validation that there is actually a unique index to support the $merge, because we assume that the default value of "on" is definitely unique.

But this doesn't account for collation, as the values of _id might not be unique, according to the providing collation, which might lead to $merge perfoming whenNotMatched action, while actually a matching document exists.

See repro.js