I could have this wrong :ermm:, but:
The statement in the article referenced may apply only to the initial snapshot generation on the publisher (ie to generation of the schema snapshot) for some configurations of merge replication; no locks are needed on published tables because the schema snapshot contains no bcp data describing table content. When during subscriber initialisation a snapshot for a subscriber in the same partition has not yet been generated, and the publication allows subscribers to initiate snapshots (which avoids the slow process of initialising the subscriber using remote select statements) a partition snapshot has to be generated to contain the BCP data for the published tables, and may require locks on the published tables (it can and presumably will use READPAST hints to avoid lock clashes that would make it wait, but that doesn't stop it taking locks). So I think the answer to the question is wrong for the case where some of the publication's tables use parametrized row filters and the publication allows subscribers to initialise snapshots.
I don't really understand this stuff - haven't read enough about merge replication. I could have it completely mixed up.