> the adjacency list is by far the most natural to meThat is why such models are sometimes called "naive trees" - it is the obvious way to fit the minimum requirements (and, of course, sometimes all the requirements) but some types of reporting on the structure are inefficient in the general case (or impossible without advanced query features like CTEs).
> subtrees can be generated in the database layer by recursive CTEs
Be careful with CTEs in postgres: they are an optimisation fence with regard to predicate pushing so for some query patterns you'll see index scans where you think it might be able to use seeks.
This may be the case in other DBs too. It isn't in SQL Server (usually), I don't know about others.