Sorry you're having troubles with this...I did say it would be fun I think 😛 The difference with this scenario and the way I've deployed it before in production is that I didn't have the updatable subscription, I just had the replication customised so that data deleted at the publisher wouldn't be deleted at the subscriber and also, as with your scenario, that any data at the subscriber would not be deleted during the snapshot process. The two-way element is a more interesting one!

To answer your question about filtering:

'out of the box' if you filter the publication, with an updatable subscription, any attempt to insert data at the subscriber that doesn't match the filter clause would generate an error, hence the need to modify the triggers. With the modification to remove the error check, only the data that matches the filter criteria would be pushed back to the publisher and anything else is ignored.