Is there a different, preferred way of writing EdgePredicates?
No. The BGL follows the style of the STL for passing predicates, which always passes by value. When the algorithm is decomposed into a bunch of functions, it has to copy the predicate (but those copies are generally optimized away).
The trick, as seem to have you've discovered on your own, is to build your predicate over references (or pointers) to any objects external to the predicate.
Is there a way to get target/source from an edge_descriptor without
the graph instance?
Not safely, but if you're keeping a reference to a graph, you won't need to.